fastaiagent 1.4.1

Build, debug, evaluate, and operate AI agents. The only SDK with fork-and-rerun Agent Replay.

FastAIAgent SDK

Build, debug, evaluate, and operate AI agents. The only SDK with Agent Replay — fork-and-rerun debugging — and a zero-ceremony Local UI that ships inside the Python wheel.

Works standalone or connected to the FastAIAgent Platform for visual editing, production monitoring, and team collaboration.

---

Quickstart

from fastaiagent import Agent, LLMClient

# Create an LLM client
llm = LLMClient(provider="openai", model="gpt-4o")

# Create an agent
agent = Agent(
    name="my-agent",
    system_prompt="You are a helpful assistant.",
    llm=llm,
)

# Run it
result = agent.run("What is the capital of France?")
print(result.output)
print(result.trace_id)  # every run is traced — use this ID for replay/debugging

Multimodal — images and PDFs as first-class inputs

from fastaiagent import Agent, LLMClient, Image, PDF

agent = Agent(name="claims", llm=LLMClient(provider="anthropic", model="claude-sonnet-4-20250514"))

result = agent.run([
    "Compare the photo to the policy and assess the claim.",
    Image.from_file("damage.jpg"),
    PDF.from_file("policy.pdf"),
])
print(result.output)

The same code works against OpenAI, Azure, Anthropic, Bedrock, and Ollama — provider-specific wire formatting (and OpenAI's tool-message workaround) is handled inside LLMClient. See docs/multimodal/.

Debug a failing agent in 30 seconds

from fastaiagent.trace import Replay

# Load a trace from a production failure
replay = Replay.load("trace_abc123")

# Step through to find the problem
replay.step_through()
# Step 3: LLM hallucinated the refund policy ← found it

# Fork at the failing step, fix, rerun
forked = replay.fork_at(step=3)
forked.modify_prompt("Always cite the exact policy section...")
result = forked.rerun()

No other SDK can do this.

Pause for human approval. For days.

from fastaiagent import Chain, FunctionTool, Resume, SQLiteCheckpointer, interrupt
from fastaiagent.chain.node import NodeType


def approve(amount: str):
    if int(amount) > 10_000:
        decision = interrupt(reason="manager_approval", context={"amount": int(amount)})
        return {"approved": decision.approved}
    return {"approved": True}


chain = Chain("refund-flow", checkpointer=SQLiteCheckpointer())
chain.add_node(
    "approve",
    tool=FunctionTool(name="approve_tool", fn=approve),
    type=NodeType.tool,
    input_mapping={"amount": "{{state.amount}}"},
)

from fastaiagent._internal.async_utils import run_sync

# First run — suspends and the process can exit cleanly.
result = chain.execute({"amount": 50_000}, execution_id="refund-abc")
assert result.status == "paused"

# Hours, days, or a server restart later, in any process:
result = run_sync(chain.aresume(
    "refund-abc",
    resume_value=Resume(approved=True, metadata={"approver": "alice"}),
))
assert result.status == "completed"

Crash-proof agents (real SIGKILL resumes at the last checkpoint), SQLite locally / Postgres in production (same Protocol surface), the @idempotent decorator that makes charge_customer safe to call inside a paused node, and a built-in /approvals UI to drive the resume from a browser. See docs/durability/.

See every trace, eval, and prompt in your browser — no Docker, no signup

pip install 'fastaiagent[ui]'
fastaiagent ui

Opens a polished web UI at http://127.0.0.1:7842. Every agent run you execute lands here — span tree with Gantt-style timing, JSON-viewer inspector, Agent Replay fork-and-rerun in the browser, eval runs with pass-rate trend charts, prompt editor with version lineage, guardrail events, agent scorecards, and a read-only browser + search playground for every LocalKB you've built. Everything stored in one SQLite file at ./.fastaiagent/local.db. Bcrypt-hashed local auth. Nothing phones home.

See your Chain / Swarm / Supervisor topology rendered as a graph

Pass your runners to build_app(runners=[...]) to enable the interactive React Flow topology view at /workflows/{type}/{name} — agent / HITL / function nodes, conditional edges, swarm handoffs, supervisor delegation arrows, all with click-to-inspect node detail panels:

import uvicorn
from fastaiagent import Agent, Chain
from fastaiagent.ui.server import build_app

researcher = Agent(name="researcher", llm=llm)
writer     = Agent(name="writer",     llm=llm)

chain = Chain("research-then-summarise")
chain.add_node("research",  agent=researcher)
chain.add_node("summarize", agent=writer)
chain.connect("research", "summarize")

# Register the chain so the topology endpoint can render it.
app = build_app(runners=[chain])
uvicorn.run(app, host="127.0.0.1", port=7843)
# → open http://127.0.0.1:7843/workflows/chain/research-then-summarise

Without runners=[...] the trace list, agent stats, and analytics still populate from runtime spans — but /workflows/chain/<name> shows a "No topology available" callout with the registration recipe above. Same pattern works for Swarm and Supervisor. See examples/47_workflow_topology.py and docs/ui/workflow-visualization.md for the full reference.

Iterate on prompts in the browser — Prompt Playground

The Prompt Playground at /playground is the inner-loop iteration surface for prompts: pick one from the registry, fill in its {{variables}}, choose a provider/model, click Run, watch the response stream back token-by-token. Edit the template inline for one-off experiments, attach an image for vision models, then click Save as eval case to append the input/output pair to a JSONL dataset that loads directly via Dataset.from_jsonl(). Every run emits a trace tagged fastaiagent.source = "playground" so playground experiments share the same observability surface as production runs.

See docs/ui/playground.md and examples/49_prompt_playground.py for the walkthrough.

See what your agent is made of — Agent Dependency Graph

The Dependencies tab on any /agents/{name} page renders a structural graph of the agent: every tool, knowledge base, prompt, guardrail, and model appears as a node radiating out from the agent centre. Tools that the LLM has called but weren't registered show up in amber so hallucinated tool names are visible at a glance. For Supervisors every Worker appears as a sub-agent with its own subtree; for Swarms peers appear as siblings with handoff edges. Click any node to inspect its details and jump to its own page.

See docs/ui/agent-dependencies.md and examples/50_agent_dependencies.py for the walkthrough.

Debug what your guardrails did — Guardrail Event Detail

Every guardrail firing already shows up on /guardrails. Click any row to open its detail page with three panels — what triggered it, which rule matched, what happened next — plus an execution-context timeline of the surrounding spans and the other guardrails that ran on the same content. For filtered events the third panel renders a before/after diff of the rewritten content; for llm_judge rules it shows the judge prompt + response inline. A Mark as false positive button flips a flag stored on the event row so you can curate signal vs. noise without ever editing the DB — and a new FP: yes / FP: no filter on the list page hides the noise once you've marked it.

See docs/ui/guardrail-events.md and examples/51_guardrail_events.py for the walkthrough.

Compare any two traces — Trace Comparison

Generalises Replay's "original vs forked" diff to any two traces, so you can answer "why did Monday's run differ from Friday's?", A/B-test two prompts on the same input, or spot a regression after a model change. Multi-select two rows on /traces → Compare in the action bar; or use Compare with… on any trace detail page. The view shows summary delta cards (duration, tokens, cost, span count) over a span-aligned table — server-side alignment matches by name first then position, classifying each row as same / slower / faster / different output / new in A / new in B. Click any row to expand side-by-side input / output / attributes diffs powered by react-diff-viewer-continued. URL is bookmarkable: /traces/compare?a=<id>&b=<id>.

See docs/ui/trace-comparison.md and examples/52_trace_compare.py for the walkthrough.

Curate eval cases inline — Eval Dataset Editor

Datasets stay JSONL on disk (same files Dataset.from_jsonl() already loads) — but the editor at /datasets replaces the script-edit-rerun loop with point-and-click CRUD. Add, edit, duplicate, delete cases; upload images for multimodal cases (the typed-parts shape is preserved on disk so cases stay framework-runnable); import / export JSONL with line-numbered errors on bad input; and Run eval kicks off the existing eval framework against the dataset and surfaces the resulting run_id in /evals. The Playground's Save as eval case button now combos over existing datasets with a + New escape hatch so the inner-loop iteration feeds outer-loop curation without copy-paste.

See docs/ui/datasets.md and examples/53_dataset_editor.py for the walkthrough.

Find any trace — Richer Trace Filtering

The Traces filter bar is now production-grade. FTS5-backed full-text search across LLM prompts and responses (gen_ai.prompt, gen_ai.response.text, with fastaiagent.* namespaced fallbacks) — sub-second on 1k spans, regression-tested, with LIKE fallback for legacy DBs. Custom date-range picker (react-day-picker) alongside the quick ranges (15m, 1h, 24h, 7d, 30d, All). More filters disclosure with duration and cost ranges. Saved filter presets (project-scoped) — capture every active filter, name it, one-click reapply. 300 ms debounced search. And URL state: every active filter mirrors into ?key=value query params, so refresh, bookmark, share, and back/forward all preserve filter state.

See docs/ui/trace-filters.md and examples/54_trace_filters.py for the walkthrough.

Other Local UI surfaces

• Multimodal trace rendering — image thumbnails and PDF cards render inline in the trace input/output tabs, no raw base64. (docs/ui/multimodal.md)

• Checkpoint inspector at /executions/{id} — vertical timeline of every checkpoint with status, expandable state snapshots, automatic state diff between adjacent rows, and an idempotency-cache panel. (docs/ui/checkpoint-inspector.md)

• Cost tracking at the bottom of /analytics — three tabs (by model / by agent / by chain node) backed by GET /api/analytics/costs. Reuses the same pricing table the per-trace cost column uses, so the numbers always agree. (docs/ui/cost-tracking.md)

• Export trace as JSON — Export button on every trace detail page opens a dialog with Include image / PDF data and Include checkpoint state toggles. Same payload from the CLI:

  fastaiagent export-trace --trace-id <id> --output trace.json
  

  (docs/ui/export-trace.md)

• Project scoping — every record the SDK writes carries a project_id resolved from ./.fastaiagent/config.toml (created lazily on the first agent.run() from a fresh directory). Multiple projects can share one Postgres without cross-contamination; the header breadcrumb reads Local UI // <project-id> // …. (docs/ui/projects.md)

See docs/ui/ for the full tour; the KB browser is documented at docs/ui/kb.md.

Evaluate agents systematically

from fastaiagent.eval import evaluate

results = evaluate(
    agent_fn=my_agent.run,
    dataset="test_cases.jsonl",
    scorers=["correctness", "relevance"]
)
print(results.summary())
# correctness: 92% | relevance: 88%

Trace any agent — yours or LangChain/CrewAI

import fastaiagent
fastaiagent.integrations.langchain.enable()

# Your existing LangChain agent, now with full tracing
result = langchain_agent.invoke({"input": "..."})
# → Traces stored locally or pushed to FastAIAgent Platform

Build agents with guardrails and cyclic workflows

from fastaiagent import Agent, Chain, LLMClient, Guardrail
from fastaiagent.guardrail import no_pii, json_valid

agent = Agent(
    name="support-bot",
    system_prompt="You are a helpful support agent...",
    llm=LLMClient(provider="openai", model="gpt-4o"),
    tools=[search_tool, refund_tool],
    guardrails=[no_pii(), json_valid()]
)

# Chains with loops (retry until quality is good enough)
chain = Chain("support-pipeline", state_schema=SupportState)
chain.add_node("research", agent=researcher)
chain.add_node("evaluate", agent=evaluator)
chain.add_node("respond", agent=responder)
chain.connect("research", "evaluate")
chain.connect("evaluate", "research", max_iterations=3, exit_condition="quality >= 0.8")
chain.connect("evaluate", "respond", condition="quality >= 0.8")

result = chain.execute({"message": "My order is late"}, trace=True)

Deploying

A fastaiagent agent is a plain Python object — wrap it in any web framework and ship it anywhere Python runs. docs/deployment has copy-paste recipes for:

• FastAPI + Uvicorn — the baseline. Works on a laptop or any VM / container.
• Docker → Cloud Run / Fly / Render / Railway — one Dockerfile, four managed container platforms.
• Modal — serverless Python with no container work.
• Replicate (Cog) — public inference endpoint.

Every recipe exposes the same POST /run + POST /run/stream contract so callers don't care where the agent lives. See the runnable starter: examples/33_deploy_fastapi.py.

Expose agents as MCP servers (Claude Desktop / Cursor / Continue / Zed)

Any Agent or Chain becomes an MCP server with one line:

from fastaiagent import Agent, LLMClient

agent = Agent(name="research_assistant", llm=LLMClient(provider="openai", model="gpt-4o"))

if __name__ == "__main__":
    import asyncio
    asyncio.run(agent.as_mcp_server(transport="stdio").run())

Register it in claude_desktop_config.json:

{
  "mcpServers": {
    "research-assistant": {
      "command": "python",
      "args": ["/absolute/path/to/my_agent.py"]
    }
  }
}

Claude Desktop now treats your fastaiagent as a callable tool. Same config shape for Cursor / Continue / Zed. Or use the CLI: fastaiagent mcp serve my_agent.py:agent. See docs/tools/mcp-server.md.

Install: pip install 'fastaiagent[mcp-server]'.

Peer-to-peer swarms with handoffs

Beyond the central-coordinator Supervisor/Worker pattern, agents can hand off to each other directly:

from fastaiagent import Agent, LLMClient, Swarm

llm = LLMClient(provider="openai", model="gpt-4o-mini")

triage = Agent(name="triage", llm=llm, system_prompt="Hand off to the right specialist.")
coder = Agent(name="coder", llm=llm, system_prompt="Answer Python questions.")
writer = Agent(name="writer", llm=llm, system_prompt="Help with prose.")

swarm = Swarm(
    name="help_desk",
    agents=[triage, coder, writer],
    entrypoint="triage",
    handoffs={"triage": ["coder", "writer"], "coder": [], "writer": []},
)
result = swarm.run("How do I reverse a list in Python?")

The currently active agent decides when to transfer control — no central LLM. See docs/agents/swarm.md for the full guide, and Swarm vs Supervisor for when to pick which.

Long-term memory with composable blocks

Beyond a sliding window, layer static facts, a rolling summary, semantic recall, and fact extraction into one memory object:

from fastaiagent import Agent, LLMClient, ComposableMemory, AgentMemory
from fastaiagent import StaticBlock, SummaryBlock, VectorBlock, FactExtractionBlock
from fastaiagent.kb.backends.faiss import FaissVectorStore

llm = LLMClient(provider="openai", model="gpt-4o-mini")

agent = Agent(
    name="assistant",
    llm=llm,
    memory=ComposableMemory(
        blocks=[
            StaticBlock("User is Upendra. Prefers terse answers."),
            SummaryBlock(llm=llm, keep_last=10, summarize_every=5),
            VectorBlock(store=FaissVectorStore(dimension=384)),
            FactExtractionBlock(llm=llm, max_facts=100),
        ],
        primary=AgentMemory(max_messages=20),
    ),
)

VectorBlock works with any VectorStore (Qdrant / Chroma / custom). Write your own block by subclassing MemoryBlock with two methods. See docs/agents/memory.md.

Swap the KB storage layer

Default LocalKB ships with FAISS + BM25 + SQLite — zero setup. Point at Qdrant, Chroma, or your own backend with one kwarg:

from fastaiagent.kb import LocalKB
from fastaiagent.kb.backends.qdrant import QdrantVectorStore

kb = LocalKB(
    name="product-docs",
    search_type="vector",
    vector_store=QdrantVectorStore(
        url="http://localhost:6333",
        collection="product-docs",
        dimension=1536,
    ),
)
kb.add("docs/")
results = kb.search("refund policy", top_k=5)

Adapters shipped: FAISS, BM25, SQLite (defaults), Qdrant (fastaiagent[qdrant]), Chroma (fastaiagent[chroma]). Write your own against the VectorStore / KeywordStore / MetadataStore protocols — see docs/knowledge-base/backends.md.

Platform-hosted KBs. For KBs uploaded and managed on the FastAIAgent platform, use fa.PlatformKB(kb_id=...) — same .search() / .as_tool() surface, retrieval (hybrid + rerank + relevance gate) runs on the platform. See docs/knowledge-base/platform-kb.md.

Shape agent behavior with middleware

Compose pre/post model hooks and tool wrappers without subclassing Agent:

from fastaiagent import Agent, LLMClient, TrimLongMessages, RedactPII, ToolBudget

agent = Agent(
    name="controlled",
    llm=LLMClient(provider="openai", model="gpt-4o"),
    tools=[search_tool],
    middleware=[
        TrimLongMessages(keep_last=30),   # cap history size
        RedactPII(),                      # scrub emails/phones/SSNs both directions
        ToolBudget(max_calls=5),          # cooperatively stop after 5 tool calls
    ],
)

Write your own by subclassing AgentMiddleware and overriding before_model, after_model, or wrap_tool. See docs/agents/middleware.md for ordering, hook reference, and custom patterns.

Multi-agent teams with context

from fastaiagent import Agent, LLMClient, RunContext, Supervisor, Worker, tool

@tool(name="get_tickets")
def get_tickets(ctx: RunContext[AppState], status: str) -> str:
    """Get support tickets for the current user."""
    return ctx.state.db.query("tickets", user_id=ctx.state.user_id, status=status)

support = Agent(name="support", llm=llm, tools=[get_tickets], system_prompt="Handle tickets.")
billing = Agent(name="billing", llm=llm, tools=[get_billing], system_prompt="Handle billing.")

supervisor = Supervisor(
    name="customer-service",
    llm=LLMClient(provider="openai", model="gpt-4o"),
    workers=[
        Worker(agent=support, role="support", description="Manages tickets"),
        Worker(agent=billing, role="billing", description="Handles billing"),
    ],
    system_prompt=lambda ctx: f"You lead support for {ctx.state.company}. Be helpful.",
)

# Context flows to all workers and their tools
ctx = RunContext(state=AppState(db=db, user_id="u-1", company="Acme"))
result = supervisor.run("Show my open tickets and billing", context=ctx)

# Stream the supervisor's response
async for event in supervisor.astream("Help me", context=ctx):
    if isinstance(event, TextDelta):
        print(event.text, end="")

Connect to FastAIAgent Platform (optional)

import fastaiagent as fa

fa.connect(api_key="fa-...", project="my-project")

# Traces automatically sent to platform dashboard
result = agent.run("Help me")

# Pull versioned prompts from platform
prompt = PromptRegistry().get("support-prompt")

# Publish eval results to platform
results = evaluate(agent, dataset=dataset)
results.publish()

SDK works standalone. Platform adds: production observability, prompt management, evaluation dashboards, team collaboration, HITL approval workflows.

Free tier available →

---

Install

pip install fastaiagent

With optional integrations:

pip install "fastaiagent[openai]"       # OpenAI auto-tracing
pip install "fastaiagent[langchain]"    # LangChain auto-tracing
pip install "fastaiagent[kb]"           # Local knowledge base
pip install "fastaiagent[all]"          # Everything

Documentation

• Getting Started
• Agent Replay Guide
• Building Chains with Cycles
• Guardrails
• Evaluation
• API Reference

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

License

Apache 2.0 — see LICENSE.