a2a-utils 0.2.0

A collection of utilities for discovering, communicating, and authenticating with A2A Servers

A2A Utils Python

This package is a collection of utility functions that allows agents to discover, communicate, and authenticate with A2A Servers (remote agents). It does this by introudcing a number of classes (e.g. A2ASession) that solve typical problems associated with implementing the A2A Client in an LLM. For example, how should a remote agent's identity and capabilities be represented to the client agent? What should the client agent be shown from the remote agent's response? And how should the client agent handle large Artifacts that would overload the context? See the "💡 Problems and Solutions" section to learn more about the problems and solutions.

📦 Installation

To install with pip:

pip install a2a-utils

To install with uv:

uv add a2a-utils

🚀 Quick Start

from pathlib import Path

from a2a_utils import (
    AgentManager,
    A2ASession,
    JSONTaskStore,
    LocalFileStore,
    TaskForLLM,
)

async def main() -> None:
    # Add remote agents
    agent_manager: AgentManager = AgentManager({
        "weather": {
            "url": "https://weather.example.com/.well-known/agent-card.json",
        },
        "stock-tracker": {
            "url": "https://example.com/finance-agent/agent-card.json",
            "custom_headers": {"X-API-Key": "key_123"},
        },
    })

    # Initiate A2ASession
    a2a_session: A2ASession = A2ASession(
        agent_manager=agent_manager,
        # Save Tasks as JSON files
        task_store=JSONTaskStore(Path("./storage/tasks")),
        # Save Artifact files (PDFs, images, etc.)
        file_store=LocalFileStore(Path("./storage/files")),
    )

    # Send a message
    # Returns LLM-friendly TaskForLLM or MessageForLLM
    response: TaskForLLM | MessageForLLM = await a2a_session.send_message(
        "weather", "What's the forecast for Tokyo?"
    )

    # Continue the conversation
    response_2: TaskForLLM | MessageForLLM = await a2a_session.send_message(
        "weather",
        "How about Osaka?",
        context_id=response.context_id,
    )

💡 Problems and Solutions

How should a remote agent's identity be represented to the client agent?

• Agent name can't be used because two remote agents might have the same name. Agent Card URL can't be used because authentication headers sent to the agent (e.g. X-API-Key) might change the agent. Authentication headers can't be exposed to the agent for security reasons.

This package introduces AgentManager and agent IDs that stores an Agent Card URL and custom headers. The client agent can use the agent ID to send messages to the remote agent without exposing the Agent Card URL or authentication headers to it.

How should a remote agent's capabilities be represented to the client agent?

• The agent's name, description, skill names, and skill descriptions are useful to the client agent. However, showing everything at once to the client agent could overload the context. The client agent should be able to view the remote agents in more depth if they need to.

This package introduces AgentManager and get_agents_for_llm method which returns a summary of the agents at different detail levels. The client agent can view the remote agent's Agent Card in more detail with the get_agent_for_llm method.

What should the client agent be shown from the remote agent's response?

• A remote agent returns a Task or Message. A Task is a complicated object containing the Task Status, Aritfacts, a history of Task Status updates, metadata, etc. not all of which should be added to the LLM's context window. However, it's necessary to share some elements of the response with the client agent. For example, the context ID is required to continue the conversation, Artifacts are the result of the Task, etc.

This package introduces LLM-friendly types that are subsets of A2A types: TaskForLLM, MessageForLLM, TaskStatusForLLM, ArtifactForLLM, TextPartForLLM, DataPartForLLM, and FilePartForLLM.

How should the client agent handle large Artifacts that would overload the context?

• Most LLMs have a context window of 128K tokens (~512K characters). Artifacts can easily exceed this. Even if they don't exceed this, tokens increase cost and degrade LLM output quality.

This package automatically summarises Artifacts that are more than send_message_character_limit characters when JSON stringified. For example, for text Artifacts the first send_message_character_limit / 2 characters are shown, followed by [... X characters omitted ...], followed by the last send_message_character_limit / 2 characters. To view the characters that were omitted, the LLM can use the view_text_artifact method, specifying the lines to view.

How should the client agent ensure that it has access to Tasks and Artifacts if the remote agent goes offline or has a retention policy?

• Agent conversations can be continued days or weeks after they started. In that time, the remote agent might have gone offline or only keep Tasks and Artifacts for X days.

This package introduces A2ASession and a JSONTaskStore which automatically saves Task and Artifact(s) as JSON files. When the client agent uses tools like view_text_artifact, the Task Store is checked first.

How should the client agent handle files?

• Remote agents can send abitrary files such as text, documents, presentations, spreadsheets, audio, images, videos, etc. The files might be Base64 encoded or sent as a downloadable URL.

This package introduces FileStore, an abstract class similar to the TaskStore, and LocalFileStore, an implementation of FileStore that saves files locally. It is out of this package's scope to implement tools to interact with them as A2A supports sending every type of file. However, if the client agent has access to Bash commands and the files are saved locally, it should be straightforward for it to interact with them.

📖 API Reference

A2ASession

Main interface for sending messages to A2A agents and viewing artifacts.

from pathlib import Path
from a2a_utils import A2ASession, AgentManager, ArtifactSettings, JSONTaskStore, LocalFileStore

session = A2ASession(
    agent_manager=AgentManager({
        "research-bot": {"url": "https://research-bot.example.com/.well-known/agent-card.json"}
    }),
    task_store=JSONTaskStore(Path("./storage/tasks")),
    file_store=LocalFileStore(Path("./storage/files")),
)

Parameter	Type	Required	Description
agent_manager	AgentManager	Yes	The agent manager instance
task_store	TaskStore | None	No	Task store for persistence (default: InMemoryTaskStore)
file_store	FileStore | None	No	File store for saving file artifacts (default: None)
artifact_settings	ArtifactSettings | None	No	Minimization/view settings (default: ArtifactSettings())

file_store determines what FilePartForLLM shows:

FileStore	Source	uri	bytes
configured	FileWithBytes	None	{"_saved_to": ["/storage/task-123/art-789/report.pdf"]}
configured	FileWithUri	{"_saved_to": ["/storage/task-123/art-789/chart.png"]}	None
not configured	FileWithBytes	None	{"_error": "No FileStore configured. Cannot access file bytes."}
not configured	FileWithUri	"https://cdn.example.com/chart.png"	None

artifact_settings determines how Artifacts are minimized and viewed:

from a2a_utils import ArtifactSettings

settings = ArtifactSettings(
    send_message_character_limit=100_000,
    minimized_object_string_length=10_000,
    view_artifact_character_limit=100_000,
)

Field	Type	Default	Description
send_message_character_limit	int	50,000	Character limit above which artifacts are minimized in send_message
minimized_object_string_length	int	5,000	Max length for individual string values within minimized data objects
view_artifact_character_limit	int	50,000	Character limit for output from view_text_artifact / view_data_artifact

async send_message(agent_id: str, message: str, *, context_id: str | None = None, task_id: str | None = None) -> TaskForLLM | MessageForLLM

Send a message to an A2A agent. The returned task is automatically saved to the task store. Artifacts are auto-minimized, and file parts are saved via the file store.

Parameter	Type	Required	Description
agent_id	str	Yes	Registered agent identifier
message	str	Yes	The message content to send
context_id	str | None	No	Context ID to continue a conversation (auto-generated when None)
task_id	str | None	No	Task ID to attach to the message

from a2a_utils import TaskForLLM, MessageForLLM

response = await session.send_message(
    "research-bot", "Find recent papers on quantum computing"
)

Example result (TaskForLLM):

{
    "id": "task-123",
    "context_id": "ctx-456",
    "kind": "task",
    "status": {
        "state": "completed",
        "message": "I found three recent papers on quantum computing and retrieved the abstract for the most recent one."
    },
    "artifacts": [
        {
            "artifact_id": "art-789",
            "description": "Search results for quantum computing papers",
            "name": "Search Results",
            "parts": [
                {
                    "kind": "data",
                    "data": [
                        {
                          "title": "Quantum Error Correction Advances",
                          "year": 2025,
                          "authors": "Chen et al."
                        },
                        {
                          "title": "Topological Quantum Computing Survey",
                          "year": 2024,
                          "authors": "Nakamura et al."
                        },
                        {
                          "title": "Fault-Tolerant Logical Qubits",
                          "year": 2024,
                          "authors": "Wang et al."
                        }
                    ]
                }
            ]
        },
        {
            "artifact_id": "art-790",
            "description": "Abstract of 'Quantum Error Correction Advances' by Chen et al.",
            "name": "Abstract",
            "parts": [
                {
                    "kind": "text",
                    "text": "Quantum computing has seen rapid advances in error correction.\nRecent work demonstrates fault-tolerant logical qubits at scale.\nThis paper surveys progress in quantum error correction from 2023-2025.\nWe review surface codes, color codes, and novel hybrid approaches.\nKey results include a 10x reduction in logical error rates.\nThese improvements bring practical quantum computing closer to reality.\nWe also discuss remaining challenges in qubit connectivity.\nFinally, we outline a roadmap for achieving fault-tolerant quantum computation."
                }
            ]
        }
    ]
}

Continue the conversation using context_id:

response_2 = await session.send_message(
    "research-bot",
    "Summarize the most recent result",
    context_id=response.context_id,
)

Returns: TaskForLLM | MessageForLLM

async view_text_artifact(agent_id, task_id, artifact_id, *, line_start=None, line_end=None, character_start=None, character_end=None) -> ArtifactForLLM

View text content from an artifact with optional line or character range. Line selection (1-based, inclusive) and character selection (0-based, Python slice semantics) are mutually exclusive.

Parameter	Type	Required	Description
agent_id	str	Yes	Agent ID for remote artifact retrieval
task_id	str	Yes	The task containing the artifact
artifact_id	str	Yes	The artifact identifier
line_start	int | None	No	Starting line number (1-based, inclusive)
line_end	int | None	No	Ending line number (1-based, inclusive)
character_start	int | None	No	Starting character index (0-based, inclusive)
character_end	int | None	No	Ending character index (0-based, exclusive)

result = await session.view_text_artifact(
    "research-bot", "task-123", "art-790", line_start=1, line_end=3
)

Example result:

{
    "artifact_id": "art-790",
    "description": "Abstract of 'Quantum Error Correction Advances' by Chen et al.",
    "name": "Abstract",
    "parts": [
        {
            "kind": "text",
            "text": "Quantum computing has seen rapid advances in error correction.\nRecent work demonstrates fault-tolerant logical qubits at scale.\nThis paper surveys progress in quantum error correction from 2023-2025."
        }
    ]
}

Returns: ArtifactForLLM

async view_data_artifact(agent_id: str, task_id: str, artifact_id: str, *, json_path: str | None = None, rows: int | list[int] | str | None = None, columns: str | list[str] | None = None) -> ArtifactForLLM

View structured data from an artifact with optional filtering.

Parameter	Type	Required	Description
agent_id	str	Yes	Agent ID for remote artifact retrieval
task_id	str	Yes	The task containing the artifact
artifact_id	str	Yes	The artifact identifier
json_path	str | None	No	Dot-separated path to extract specific fields
rows	int | list[int] | str | None	No	Row selection (e.g. "0-10", [0, 5, 9], "all")
columns	str | list[str] | None	No	Column selection (e.g. ["name", "age"], "all")

result = await session.view_data_artifact(
    "research-bot", "task-123", "art-789",
    rows="0-1", columns=["title", "year"],
)

Example result:

{
    "artifact_id": "art-789",
    "description": "Search results for quantum computing papers",
    "name": "Search Results",
    "parts": [
        {
            "kind": "data",
            "data": [
                {"title": "Quantum Error Correction Advances", "year": 2025},
                {"title": "Topological Quantum Computing Survey", "year": 2024}
            ]
        }
    ]
}

Returns: ArtifactForLLM

🤖 AgentManager

Manages A2A agent cards keyed by user-defined agent IDs.

from a2a_utils import AgentManager

# From dict
manager = AgentManager({
    "language-translator": {
        "url": "https://example.com/language-translator/agent-card.json",
        "custom_headers": {"Authorization": "Bearer tok_123"},
    }
})

# From JSON file
manager = AgentManager("./agents.json")

# Empty — add agents later
manager = AgentManager()

async add_agent(agent_id: str, url: str, custom_headers: dict[str, str] | None = None) -> None

Register a new agent at runtime.

Parameter	Type	Required	Description
agent_id	str	Yes	User-defined agent identifier
url	str	Yes	Agent card URL
custom_headers	dict[str, str] | None	No	Custom HTTP headers

Raises: ValueError if agent_id is already registered.

await manager.add_agent(
    "code-reviewer",
    "https://review.example.com/.well-known/agent-card.json",
    custom_headers={"X-API-Key": "key_123"},
)

async get_agent(agent_id: str) -> AgentURLAndCustomHeaders | None

Retrieve agent by ID. Note: this should NOT be added to the LLM's context, use get_agent_for_llm instead.

Parameter	Type	Required	Description
agent_id	str	Yes	User-defined agent identifier

Returns: AgentURLAndCustomHeaders | None

agent = await manager.get_agent("language-translator")

Example result:

AgentURLAndCustomHeaders(
    agent_card=AgentCard(
        name="Universal Translator",
        description="Translate text and audio between 50+ languages",
        url="https://translate.example.com",
        version="1.0.0",
        capabilities=AgentCapabilities(streaming=False, pushNotifications=False),
        skills=[
            AgentSkill(
                id="translate-text",
                name="Translate Text",
                description="Translate text between any supported language pair",
                tags=["translate", "text", "language"],
                examples=["Translate 'hello' to French"],
            ),
            AgentSkill(
                id="translate-audio",
                name="Translate Audio",
                description="Translate audio between any supported language pair",
                tags=["translate", "audio", "language"],
            )
        ],
        defaultInputModes=["text", "audio/mpeg"],
        defaultOutputModes=["text", "audio/mpeg"],
    ),
    custom_headers={"Authorization": "Bearer tok_123"},
)

Returns None if the agent ID is not registered.

async get_agents() -> dict[str, AgentURLAndCustomHeaders]

Get all registered agents. Note: this should NOT be added to the LLM's context, use get_agents_for_llm instead.

Returns: dict[str, AgentURLAndCustomHeaders]

agents = await manager.get_agents()

Example result:

{
    "language-translator": AgentURLAndCustomHeaders(
        agent_card=AgentCard(
            name="Universal Translator",
            description="Translate text and audio between 50+ languages",
            ...
        ),
        custom_headers={"Authorization": "Bearer tok_123"},
    ),
    "code-reviewer": AgentURLAndCustomHeaders(
        agent_card=AgentCard(
            name="Code Reviewer",
            description="Review code for best practices",
            ...
        ),
        custom_headers={"X-API-Key": "key_123"},
    ),
}

async get_agent_for_llm(agent_id: str, detail: str = "basic") -> dict[str, Any] | None

Generate summary for a single agent.

Parameter	Type	Required	Description
agent_id	str	Yes	User-defined agent identifier
detail	str	No	Detail level: "name", "basic" (default), "skills", or "full"

Returns: dict[str, Any] | None — summary dict or None if not found.

summary = await manager.get_agent_for_llm("language-translator")

{
  "name": "Universal Translator",
  "description": "Translate text and audio between 50+ languages"
}

async get_agents_for_llm(detail: str = "basic") -> dict[str, dict[str, Any]]

Generate summary of all agents, sorted by agent_id.

Parameter	Type	Required	Description
detail	str	No	Detail level: "name", "basic" (default), "skills", or "full"

Returns: dict[str, dict[str, Any]]

"name":

summaries = await manager.get_agents_for_llm("name")

{
  "code-reviewer": {"name": "Code Reviewer"},
  "language-translator": {"name": "Universal Translator"}
}

"basic" (default):

summaries = await manager.get_agents_for_llm()

{
  "code-reviewer": {
    "name": "Code Reviewer",
    "description": "Review code for best practices"
  },
  "language-translator": {
    "name": "Universal Translator",
    "description": "Translate text and audio between 50+ languages"
  }
}

"skills":

summaries = await manager.get_agents_for_llm("skills")

{
  "code-reviewer": {
    "name": "Code Reviewer",
    "description": "Review code for best practices",
    "skills": ["Review Code"]
  },
  "language-translator": {
    "name": "Universal Translator",
    "description": "Translate text between 50+ languages",
    "skills": ["Translate Text", "Translate Audio"]
  }
}

"full":

summaries = await manager.get_agents_for_llm("full")

{
  "code-reviewer": {
    "name": "Code Reviewer",
    "description": "Review code for best practices",
    "skills": [
      {
        "name": "Review Code",
        "description": "Review code for best practices, identify bugs, and suggest improvements"
      }
    ]
  },
  "language-translator": {
    "name": "Universal Translator",
    "description": "Translate text between 50+ languages",
    "skills": [
      {
        "name": "Translate Text",
        "description": "Translate text between any supported language pair"
      },
      {
        "name": "Translate Audio",
        "description": "Translate audio between any supported language pair"
      }
    ]
  }
}

💾 JSONTaskStore

Persists Task objects as individual JSON files. Implements the A2A SDK TaskStore ABC.

from pathlib import Path
from a2a_utils import JSONTaskStore

task_store = JSONTaskStore(Path("./storage/tasks"))

async save(task: Task) -> None

Save a task to disk. Creates ./storage/tasks/{task_id}.json.

async get(task_id: str) -> Task | None

Load a task from disk.

Returns None if the task file does not exist.

async delete(task_id: str) -> None

Delete a task from disk.

📁 Files

FileStore

Abstract base class for file storage. Implement this to use custom storage backends (S3, GCS, etc.).

from abc import ABC, abstractmethod
from a2a.types import Artifact

class FileStore(ABC):
    @abstractmethod
    async def save(self, task_id: str, artifact: Artifact) -> list[str]:
        """Save file parts from an artifact. Returns list of storage locations."""

    @abstractmethod
    async def get(self, task_id: str, artifact_id: str) -> list[str]:
        """Get storage locations for a saved artifact's files. Returns empty list if not found."""

    @abstractmethod
    async def delete(self, task_id: str, artifact_id: str) -> None:
        """Delete saved files for an artifact."""

LocalFileStore

Saves artifact file parts to the local filesystem. Files are stored at storage_dir/task_id/artifact_id/filename.

from pathlib import Path
from a2a_utils import LocalFileStore

file_store = LocalFileStore(Path("./storage/files"))

async save(task_id: str, artifact: Artifact) -> list[str]

Save file parts from an artifact to disk.

saved_paths = await file_store.save("task-123", artifact)

Example result:

["./storage/files/task-123/art-789/quarterly_report.pdf"]

async get(task_id: str, artifact_id: str) -> list[str]

Get saved file paths for an artifact.

paths = await file_store.get("task-123", "art-789")

Example result:

["./storage/files/task-123/art-789/quarterly_report.pdf"]

Returns an empty list if no files are found.

async delete(task_id: str, artifact_id: str) -> None

Delete saved files for an artifact.

await file_store.delete("task-123", "art-789")

🎨 Artifacts

The A2ASession uses the TextArtifacts and DataArtifacts classes to automatically minimize Artifacts that are returned from send_message and view Artifacts using view_text_artifact and view_data_artifact. They can also be used independently on raw data.

TextArtifacts

TextArtifacts.view(text, *, line_start=None, line_end=None, character_start=None, character_end=None, character_limit=50_000) -> str

View text content with optional line or character range selection. Supports line selection (1-based, inclusive) or character selection (0-based, Python slice semantics). These are mutually exclusive — providing both raises ValueError.

Parameter	Type	Required	Description
text	str	Yes	The text to view
line_start	int | None	No	Starting line number (1-based, inclusive)
line_end	int | None	No	Ending line number (1-based, inclusive)
character_start	int | None	No	Starting character index (0-based, inclusive)
character_end	int | None	No	Ending character index (0-based, exclusive)
character_limit	int	No	Maximum output size (default: 50,000)

Returns: str

Line selection:

from a2a_utils import TextArtifacts

text = "[INFO] Server started\n[INFO] Connected to DB\n[WARN] Cache miss\n[INFO] Request OK"
TextArtifacts.view(text, line_start=1, line_end=2)

Example result:

"[INFO] Server started\n[INFO] Connected to DB"

Character selection:

TextArtifacts.view("Hello, World!", character_start=0, character_end=5)

Example result:

"Hello"

TextArtifacts.minimize(text, *, character_limit=50_000, tip=None) -> dict[str, Any]

Minimize text content for display. If text is within the character limit, returns it unchanged. If over the limit, shows first and last halves with metadata.

Parameter	Type	Required	Description
text	str	Yes	The text content to minimize
character_limit	int	No	Character limit (default: 50,000)
tip	str | None	No	Tip string (default: None; pass a string to include one)

Returns: dict[str, Any]

Short text (under limit):

from a2a_utils import TextArtifacts

TextArtifacts.minimize("Hello, world!")

{"text": "Hello, world!"}

Long text (over limit):

from a2a_utils import TextArtifacts

TextArtifacts.minimize("x" * 60_000)

{
    "text": "xxxxxxx...\n\n[... 10,000 characters omitted ...]\n\nxxxxxxx...",
    "_total_lines": 1,
    "_total_characters": 60000,
    "_start_line_range": "1-1",
    "_end_line_range": "1-1",
    "_start_character_range": "0-25000",
    "_end_character_range": "35000-60000"
}

DataArtifacts

DataArtifacts.view(data, *, json_path=None, rows=None, columns=None, character_limit=50_000) -> Any

View structured data with optional filtering. Navigate with json_path, then filter with rows/columns.

Parameter	Type	Required	Description
data	Any	Yes	The data to view
json_path	str | None	No	Dot-separated path to extract specific fields
rows	int | list[int] | str | None	No	Row selection
columns	str | list[str] | None	No	Column selection
character_limit	int	No	Maximum output size (default: 50,000)

Returns: Any (filtered data)

from a2a_utils import DataArtifacts

data = {
    "employees": [
        {"name": "Alice", "department": "Engineering", "level": 5},
        {"name": "Bob", "department": "Design", "level": 3},
        {"name": "Carol", "department": "Engineering", "level": 4},
    ]
}

DataArtifacts.view(data, json_path="employees", rows="0-2", columns=["name", "department"])

Example result:

[
    {"name": "Alice", "department": "Engineering"},
    {"name": "Bob", "department": "Design"}
]

DataArtifacts.minimize(data, *, character_limit=50_000, minimized_object_string_length=5_000, tip=None) -> dict[str, Any]

Minimize data content for display based on type. Automatically selects the best strategy: list-of-objects gets a table summary, dicts get string truncation, strings delegate to TextArtifacts.minimize.

Parameter	Type	Required	Description
data	Any	Yes	The data to minimize
character_limit	int	No	Character limit (default: 50,000)
minimized_object_string_length	int	No	Max string length in objects (default: 5,000)
tip	str | None	No	Tip string (default: None; pass a string to include one)

Returns: dict[str, Any]

from a2a_utils import DataArtifacts

data = {
    "title": "Quarterly Report Q4 2025",
    "summary": "x" * 10_000,
    "metrics": {"revenue": 1_250_000, "growth": 12.5},
    "employees": [
        {
            "name": f"Employee {i}",
            "department": ["Eng", "Marketing", "Design", "Sales"][i % 4],
            "salary": 60_000 + i * 500,
        }
        for i in range(100)
    ],
    "tags": ["finance", "quarterly", "internal"],
}
DataArtifacts.minimize(data, character_limit=100, minimized_object_string_length=10)

{
    "data": {
        "title": "Quarterly ... [14 more chars]",
        "summary": "xxxxxxxxxx... [9,990 more chars]",
        "metrics": {
            "revenue": 1250000,
            "growth": 12.5
        },
        "employees": {
            "_total_rows": 100,
            "_columns": [
                {
                    "count": 100,
                    "unique_count": 100,
                    "types": [{
                        "name": "string",
                        "count": 100,
                        "percentage": 100.0,
                        "sample_value": "Employee 42",
                        "length_minimum": 10,
                        "length_maximum": 11,
                        "length_average": 10.9,
                        "length_stdev": 0.3
                    }],
                    "name": "name"
                },
                {
                    "count": 100,
                    "unique_count": 4,
                    "types": [{
                        "name": "string",
                        "count": 100,
                        "percentage": 100.0,
                        "sample_value": "Engineering",
                        "length_minimum": 5,
                        "length_maximum": 11,
                        "length_average": 7.75,
                        "length_stdev": 2.4
                    }],
                    "name": "department"
                },
                {
                    "count": 100,
                    "unique_count": 100,
                    "types": [{
                        "name": "int",
                        "count": 100,
                        "percentage": 100.0,
                        "sample_value": 75000,
                        "minimum": 60000,
                        "maximum": 109500,
                        "average": 84750,
                        "stdev": 14505.75
                    }],
                    "name": "salary"
                }
            ],
            "_json_path": "employees"
        },
        "tags": ["finance", "quarterly", "internal"]
    }
}

Lists of dictionaries are summarized as table summaries (see summarize_table) and lists of values as value summaries (see summarize_values).

DataArtifacts.summarize_table(data) -> list[dict[str, Any]]

Generate a summary of tabular data (list of dicts). Returns one summary dict per column with count, unique count, and per-type statistics.

Parameter	Type	Required	Description
data	list[dict[str, Any]]	Yes	Table rows

Returns: list[dict[str, Any]]

from a2a_utils import DataArtifacts

# Same dataset as the minimize example above
data = [
    {
        "name": f"Employee {i}",
        "department": ["Eng", "Marketing", "Design", "Sales"][i % 4],
        "salary": 60_000 + i * 500,
    }
    for i in range(100)
]

DataArtifacts.summarize_table(data)

[
    {
        "count": 100,
        "unique_count": 100,
        "types": [
            {
                "name": "string",
                "count": 100,
                "percentage": 100.0,
                "sample_value": "Employee 42",
                "length_minimum": 10,
                "length_maximum": 11,
                "length_average": 10.9,
                "length_stdev": 0.3
            }
        ],
        "name": "name"
    },
    {
        "count": 100,
        "unique_count": 4,
        "types": [
            {
                "name": "string",
                "count": 100,
                "percentage": 100.0,
                "sample_value": "Engineering",
                "length_minimum": 5,
                "length_maximum": 11,
                "length_average": 7.75,
                "length_stdev": 2.4
            }
        ],
        "name": "department"
    },
    {
        "count": 100,
        "unique_count": 100,
        "types": [
            {
                "name": "int",
                "count": 100,
                "percentage": 100.0,
                "sample_value": 75000,
                "minimum": 60000,
                "maximum": 109500,
                "average": 84750,
                "stdev": 14505.75
            }
        ],
        "name": "salary"
    }
]

DataArtifacts.summarize_values(values) -> dict[str, Any] | list[Any]

Generate statistics for a list of values (like a single column). Includes count, unique count, and per-type statistics (min/max/avg/stdev for numbers, length stats for strings, etc.). If the summary would be larger than the original values, the original list is returned instead (inflation guard).

Parameter	Type	Required	Description
values	list[Any]	Yes	Values to summarize

Returns: dict[str, Any] | list[Any]

from a2a_utils import DataArtifacts

salaries = [
    95000, 72000, 105000, 68000, 88000,
    # ... ~100 salary values total, with some nulls
    None, 115000, 92000, None, 78000,
]

DataArtifacts.summarize_values(salaries)

{
    "count": 100,
    "unique_count": 87,
    "types": [
        {
            "name": "int",
            "count": 92,
            "percentage": 92.0,
            "sample_value": 95000,
            "minimum": 45000,
            "maximum": 185000,
            "average": 87250.5,
            "stdev": 28430.12
        },
        {
            "name": "null",
            "count": 8,
            "percentage": 8.0,
            "sample_value": null
        }
    ]
}

minimize_artifacts(artifacts, *, character_limit=50_000, minimized_object_string_length=5_000, saved_file_paths=None, text_tip=None, data_tip=None) -> list[ArtifactForLLM]

Minimize a list of artifacts for LLM display. Called automatically by send_message. Combines all TextParts within each artifact into a single TextPartForLLM. Handles FileParts by including file metadata and saved paths.

Parameter	Type	Required	Description
artifacts	list[Artifact]	Yes	List of artifacts to minimize
character_limit	int	No	Character limit (default: 50,000)
minimized_object_string_length	int	No	Max string length in objects (default: 5,000)
saved_file_paths	dict[str, list[str]] | None	No	Mapping of artifact_id to saved file paths
text_tip	str | None	No	Tip string for minimized text artifacts (default: None)
data_tip	str | None	No	Tip string for minimized data artifacts (default: None)

Returns: list[ArtifactForLLM]

from a2a.types import Artifact, TextPart, DataPart, FilePart, FileWithBytes
from a2a_utils import minimize_artifacts

artifacts = [
    Artifact(
        artifactId="art-123",
        description="Full text of a research paper abstract spanning several pages.",
        name="Research Paper Abstract",
        parts=[TextPart(kind="text", text="x" * 60_000)],
    ),
    Artifact(
        artifactId="art-456",
        description="Company employee directory with names, departments, and salaries.",
        name="Employee Directory",
        parts=[DataPart(kind="data", data=[
            {
                "name": f"Employee {i}",
                "department": ["Eng", "Marketing", "Design", "Sales"][i % 4],
                "salary": 60_000 + i * 500,
            }
            for i in range(100)
        ])],
    ),
    Artifact(
        artifactId="art-789",
        description="Generated quarterly financial report in PDF format.",
        name="Quarterly Report",
        parts=[
            FilePart(
                kind="file",
                file=FileWithBytes(
                    name="q4-report.pdf",
                    mimeType="application/pdf",
                    bytes="base64encodeddata...",
                ),
            ),
        ],
    ),
]

minimized = minimize_artifacts(
    artifacts,
    text_tip="Text was minimized. Call view_text_artifact() to see specific line ranges.",
    data_tip="Data was minimized. Call view_data_artifact() to navigate to specific data.",
    saved_file_paths={"art-789": ["./storage/files/task-123/art-789/q4-report.pdf"]},
)

Example result:

[
    ArtifactForLLM(
        artifact_id="art-123",
        description="Full text of a research paper abstract spanning several pages.",
        name="Research Paper Abstract",
        parts=[
            TextPartForLLM(
                kind="text",
                text="xxxxxxx...\n\n[... 10,000 characters omitted ...]\n\nxxxxxxx...",
            ),
        ],
    ),
    ArtifactForLLM(
        artifact_id="art-456",
        description="Company employee directory with names, departments, and salaries.",
        name="Employee Directory",
        parts=[
            DataPartForLLM(
                kind="data",
                data={
                    "data": {
                        "_total_rows": 100,
                        "_columns": ["..."],
                        "_tip": "Data was minimized. Call view_data_artifact() to navigate to specific data.",
                    }
                },
            ),
        ],
    ),
    ArtifactForLLM(
        artifact_id="art-789",
        description="Generated quarterly financial report in PDF format.",
        name="Quarterly Report",
        parts=[
            FilePartForLLM(
                kind="file",
                name="q4-report.pdf",
                mime_type="application/pdf",
                uri=None,
                bytes={"_saved_to": ["./storage/files/task-123/art-789/q4-report.pdf"]},
            ),
        ],
    ),
]

📋 Types

All types are frozen dataclasses exported from a2a_utils.

AgentURLAndCustomHeaders

Returned by AgentManager.get_agent() and AgentManager.get_agents().

AgentURLAndCustomHeaders(
    agent_card=AgentCard(
        name="Universal Translator",
        description="Translate text and audio between 50+ languages",
        url="https://translate.example.com",
        version="1.0.0",
        capabilities=AgentCapabilities(streaming=False, pushNotifications=False),
        skills=[
            AgentSkill(
                id="translate-text",
                name="Translate Text",
                description="Translate text between any supported language pair",
                tags=["translate", "text", "language"],
                examples=["Translate 'hello' to French"],
            ),
            AgentSkill(
                id="translate-audio",
                name="Translate Audio",
                description="Translate audio between any supported language pair",
                tags=["translate", "audio", "language"],
            )
        ],
        defaultInputModes=["text", "audio/mpeg"],
        defaultOutputModes=["text", "audio/mpeg"],
    ),
    custom_headers={"Authorization": "Bearer tok_123"},
)

Field	Type
agent_card	AgentCard
custom_headers	dict[str, str]

TaskForLLM

Returned by A2ASession.send_message() for task responses.

TaskForLLM(
    id="task-123",
    context_id="ctx-456",
    kind="task",
    status=TaskStatusForLLM(
        state=TaskState.completed,
        message=MessageForLLM(
            context_id="ctx-456",
            kind="message",
            parts=[
                TextPartForLLM(
                    kind="text",
                    text="I found three recent papers on quantum computing and retrieved the abstract for the most recent one.",
                ),
            ],
        ),
    ),
    artifacts=[
        ArtifactForLLM(
            artifact_id="art-789",
            description="Search results for quantum computing papers",
            name="Search Results",
            parts=[
                DataPartForLLM(kind="data", data=[
                    {
                        "title": "Quantum Error Correction Advances",
                        "year": 2025,
                        "authors": "Chen et al.",
                    },
                    {
                        "title": "Topological Quantum Computing Survey",
                        "year": 2024,
                        "authors": "Nakamura et al.",
                    },
                    {
                        "title": "Fault-Tolerant Logical Qubits",
                        "year": 2024,
                        "authors": "Wang et al.",
                    },
                ]),
            ],
        ),
        ArtifactForLLM(
            artifact_id="art-790",
            description="Abstract of 'Quantum Error Correction Advances' by Chen et al.",
            name="Abstract",
            parts=[
                TextPartForLLM(
                    kind="text",
                    text="Quantum computing has seen rapid advances in error correction.\nRecent work demonstrates fault-tolerant logical qubits at scale.\nThis paper surveys progress in quantum error correction from 2023-2025.\nWe review surface codes, color codes, and novel hybrid approaches.\nKey results include a 10x reduction in logical error rates.\nThese improvements bring practical quantum computing closer to reality.\nWe also discuss remaining challenges in qubit connectivity.\nFinally, we outline a roadmap for achieving fault-tolerant quantum computation.",
                ),
            ],
        ),
    ],
)

Field	Type
id	str
context_id	str
kind	str ("task")
status	TaskStatusForLLM
artifacts	list[ArtifactForLLM]

MessageForLLM

Returned by A2ASession.send_message() for message-only responses, or as TaskStatusForLLM.message.

MessageForLLM(
    context_id="ctx-456",
    kind="message",
    parts=[
        TextPartForLLM(
            kind="text",
            text="I found three recent papers on quantum computing and retrieved the abstract for the most recent one.",
        ),
    ],
)

Field	Type
context_id	str | None
kind	str ("message")
parts	list[TextPartForLLM | DataPartForLLM | FilePartForLLM]

TaskStatusForLLM

TaskStatusForLLM(
    state=TaskState.completed,
    message=MessageForLLM(
        context_id="ctx-456",
        kind="message",
        parts=[
            TextPartForLLM(
                kind="text",
                text="I found three recent papers on quantum computing and retrieved the abstract for the most recent one.",
            ),
        ],
    ),
)

Field	Type
state	TaskState
message	MessageForLLM | None

ArtifactForLLM

Returned by view_text_artifact(), view_data_artifact(), and minimize_artifacts(). Used in TaskForLLM.artifacts.

ArtifactForLLM(
    artifact_id="art-790",
    description="Abstract of 'Quantum Error Correction Advances' by Chen et al.",
    name="Abstract",
    parts=[
        TextPartForLLM(
            kind="text",
            text="Quantum computing has seen rapid advances in error correction.\nRecent work demonstrates fault-tolerant logical qubits at scale.\nThis paper surveys progress in quantum error correction from 2023-2025.\nWe review surface codes, color codes, and novel hybrid approaches.\nKey results include a 10x reduction in logical error rates.\nThese improvements bring practical quantum computing closer to reality.\nWe also discuss remaining challenges in qubit connectivity.\nFinally, we outline a roadmap for achieving fault-tolerant quantum computation.",
        ),
    ],
)

Field	Type
artifact_id	str
description	str | None
name	str | None
parts	list[TextPartForLLM | DataPartForLLM | FilePartForLLM]

TextPartForLLM

TextPartForLLM(
    kind="text",
    text="Quantum computing has seen rapid advances in error correction.\nRecent work demonstrates fault-tolerant logical qubits at scale.\nThis paper surveys progress in quantum error correction from 2023-2025.",
)

Field	Type
kind	str ("text")
text	str

DataPartForLLM

DataPartForLLM(kind="data", data=[
    {
        "title": "Quantum Error Correction Advances",
        "year": 2025,
        "authors": "Chen et al.",
    },
    {
        "title": "Topological Quantum Computing Survey",
        "year": 2024,
        "authors": "Nakamura et al.",
    },
    {
        "title": "Fault-Tolerant Logical Qubits",
        "year": 2024,
        "authors": "Wang et al.",
    },
])

Field	Type
kind	str ("data")
data	Any

FilePartForLLM

Represents a file part in artifacts and messages. uri and bytes are mutually exclusive — at most one is set.

FilePartForLLM(
    kind="file",
    name="q4-report.pdf",
    mime_type="application/pdf",
    uri=None,
    bytes={
        "_saved_to": [
            "./storage/files/task-123/art-789/q4-report.pdf",
        ],
    },
)

Field	Type	Description
kind	str ("file")	Always "file"
name	str | None	Filename from the original FilePart
mime_type	str | None	MIME type from the original FilePart
uri	str | dict[str, Any] | None	Raw URI (no FileStore) or {"_saved_to": [...]} (FileStore saved it)
bytes	dict[str, Any] | None	{"_saved_to": [...]} (FileStore saved it) or {"_error": "..."} (no FileStore)

📄 License

a2a-utils is distributed under the terms of the Apache-2.0 license.

🤝 Join the A2A Net Community

A2A Net is a site to find and share AI agents and open-source community.

• 🌍 Site: A2A Net
• 🤖 Discord: Join the Discord