fifty-agent-sdk

fifty-agent-sdk is a reusable agent loop for python. it implements a custom reACT loop with json-mode tool calls, an mcp client, and pluggable llm, state, and tool backends. it exists because the loop, the parser, the safety checks, and the runner kept getting rewritten per project. this is that loop, factored out once: write the tools, hand them to the runner, let it iterate.

At a glance

• talks to any openai-compatible chat-completions endpoint by swapping one base_url: openai, google distributed cloud, a local oss server.
• llm clients, state stores, and tools are pluggable behind protocols: bring your own, the loop stays the same.
• the run emits a typed event stream the caller consumes, so you watch the react loop step by step.
• an iteration cap and per-tool timeouts bound every run, with a fallback answer on error or cap: a loop that can't end is a loop that doesn't ship.
• zero-infra by default: no db, no redis, until you opt into an extra.

Installation

pip install fifty-agent-sdk

Optional extras:

• pip install 'fifty-agent-sdk[sql]' — enables SqlStateStore, SqlAuditSink, SQLAlchemy
• pip install 'fifty-agent-sdk[redis]' — enables RedisStateStore

Importing fifty_agent_sdk pulls neither extra; the extra symbols are re-exported lazily, and first access without the relevant extra installed raises a clear ImportError. The sql extra installs SQLAlchemy but not a database driver — bring your own async driver (e.g. aiosqlite for SQLite, asyncpg for PostgreSQL).

Requires Python >=3.11.

Quickstart

the example builds a tool, hands it to the AgentRunner, and consumes the typed event stream the run emits.

import asyncio
from typing import Any

from fifty_agent_sdk import (
    JSON_MODE_OUTPUT_FORMAT,
    AgentLoop,
    AgentRunner,
    JsonModeParser,
    MemoryStateStore,
    OpenAICompatibleClient,
    PromptSections,
    Registry,
    SafetyConfig,
    tool,
)


@tool()
async def get_weather(city: str) -> dict[str, Any]:
    """Return the current weather for a city."""
    return {"city": city, "temp_c": 21}


async def main() -> None:
    # 1. An LLM client — points at any OpenAI-compatible endpoint.
    #    Pass base_url=... to target GDC or a local OSS server instead of OpenAI.
    llm = OpenAICompatibleClient(api_key="sk-...")

    # 2. A tool registry — register the decorated tool.
    registry = Registry()
    registry.register(get_weather)

    # 3. The ReACT loop — LLM + registry + parser + prompts + safety.
    #    `output_format` shows the model the JSON envelope the parser
    #    expects; without it JsonModeParser raises ParserError on every turn.
    loop = AgentLoop(
        llm=llm,
        registry=registry,
        parser=JsonModeParser(),
        prompts=PromptSections(persona="You are helpful."),
        safety=SafetyConfig(),
        model="gpt-4o",
        output_format=JSON_MODE_OUTPUT_FORMAT,
    )

    # 4. The runner — wraps the loop with conversation-state persistence.
    runner = AgentRunner(
        loop=loop,
        state=MemoryStateStore(),
        system_prompt="You are a helpful weather assistant.",
    )

    # 5. Drive a turn and consume the event stream.
    async for event in runner.run("session-1", "What's the weather in Paris?"):
        print(event)


asyncio.run(main())

Core concepts

tools

the registry of functions the agent can call. each tool is a side-effecting action exposed to the loop, so the model can do something in the world and not just talk about it.

llm

the llm client. a protocol plus an openai-compatible adapter, so the loop talks to any chat-completions endpoint by changing one base_url. a max_tokens cap is sent as max_completion_tokens for gpt-5.x and o-series models, which reject max_tokens; max_tokens_param= overrides that choice per client.

state

the state stores. where conversation state persists between turns, with branching built in: fork a session, switch between branches, truncate back to an earlier point. MemoryStateStore needs no infrastructure, but it is process-local and non-durable: by default it lazily expires whole sessions when monotonic inactivity reaches 3,600 seconds and retains at most 1,000 sessions using LRU eviction. successful reads refresh inactivity. SqlStateStore and RedisStateStore are durable backends behind the extras.

a runner hands back the store it was built with as runner.state, so the branching calls above are reachable from a runner you already have:

store = MemoryStateStore()
runner = AgentRunner(loop=..., state=store)

runner.state is store  # True — the exact instance, never a copy or a wrapper
branch = await runner.state.fork(session_id, from_sequence=4)

configure either in-memory bound independently when an ephemeral workload needs different limits:

store = MemoryStateStore(ttl_seconds=900, max_sessions=250)

the former unbounded behavior remains available as an explicit opt-in with MemoryStateStore(ttl_seconds=None, max_sessions=None). prefer a durable backend instead when conversation state must survive process restarts.

identity is the point rather than convenience: a second store constructed over the same engine carries its own lock registry, so two writers could interleave on one session. sharing runner.state shares the serialization too.

it is read-only, for correctness and not for style. run() appends the user message, drives the loop, then appends the assistant message — a store swapped in between those appends would split one turn across two backends. assignment raises AttributeError, and mypy rejects it statically. to use a different store, construct another runner; __init__ does no i/o. the declared type is the StateStore protocol, so keep your own concretely-typed reference if you need backend-specific api like SqlStateStore.aclose().

runner.state is the supported way in. _state is private, carries no semver protection, and may be renamed or removed in a patch release.

streaming

a typed event stream the caller consumes while the loop runs. each step in the run surfaces as an event instead of waiting for a final blob.

safety

the caps that bound a run: a max-iteration ceiling on react cycles and a per-tool timeout, plus the fallback answer returned when a run errors or hits the cap. a loop that can't end is a loop that doesn't ship.

audit

the audit sinks and observability hooks. they record what the agent did, so a run can be read back after it finishes.

mcp

an mcp client over streamable http, adapted into the same registry the in-proc tools live in. a tools/call that comes back isError=True is a recoverable observation the model can reason about, not a dead run — and on_tool_error is the seam for screening that server-controlled text before the model reads it.

def screen(message: str, content: list[dict]) -> str:
    # `message` is the sdk's bounded default; `content` is the server's raw
    # error blocks (read-only). return the string the model should see.
    if any("PII" in str(block) for block in content):  # your own predicate
        return "the upstream tool failed"
    return message


client = MCPClient(MCPClientConfig(base_url=...), auth=..., on_tool_error=screen)
provider = MCPProvider(client)
await provider.attach(registry)

the hook may be sync or async, and it only ever fires on a per-call isError result — never on success, never on a transport failure (that still raises MCPError). if it raises, returns a non-string, or returns a blank string, the sdk falls back to its own bounded message and logs a warning; it can never change is_error or output.

Architecture

fifty_agent_sdk  —  module graph (from src/fifty_agent_sdk/, ground-truth imports)

src/fifty_agent_sdk/
├─ ▢ audit
├─ errors
├─ ▢ llm
├─ loop
├─ ▢ mcp
├─ ▢ observability
├─ ▢ parser
├─ prompts
├─ ▶ runner
├─ safety
├─ ▢ state
├─ streaming
└─ ▢ tools

depends (→):
   audit → errors
   llm → errors
   loop → errors
   loop → llm
   loop → observability
   loop → parser
   loop → prompts
   loop → safety
   loop → streaming
   loop → tools
   mcp → errors
   observability → llm
   parser → errors
   parser → llm
   runner → audit
   runner → errors
   runner → llm
   runner → loop
   runner → observability
   runner → state
   runner → streaming
   state → errors
   state → llm
   streaming → tools
   tools → errors
   tools → llm
   tools → mcp

legend: ▶ entry   ▢ package   name module   → depends

Highlights

• branching — first-class conversation branching on StateStore: fork, list_branches, switch_branch, branch-scoped get_messages(..., branch_id=...), plus BranchInfo and TRUNK_BRANCH_ID. a session is now a tree of branches with an active head, and append writes to the active branch (the edit-a-message / regenerate model). implemented across memory, SQL, and Redis backends, data-additive and zero-migration: existing sessions read as the trunk branch. breaking for custom StateStore implementations: they must add the new methods.
• StateStore.truncate_after(session_id, sequence, *, branch_id=None) — a destructive hard-delete of a branch's tail (messages with sequence > N), for redaction, retention, and rollback. only the target branch's own messages are removed (a fork's inherited prefix is never touched), and it is idempotent: a no-op on an unknown session or branch.

editing a turn is a consumer-side fork-then-append, and the original line stays reachable:

# Edit a turn = fork the history before it, switch onto the new branch, then
# append the edited message. `store` is any StateStore; import `ChatMessage`
# from fifty_agent_sdk.
branch = await store.fork(session_id, from_sequence=4)  # keep messages 1..4
await store.switch_branch(session_id, branch)
await store.append(session_id, ChatMessage(role="user", content="...edited..."))
await store.get_messages(session_id, branch_id="trunk")  # original line intact

Links

• Homepage
• Repository
• Issues
• Changelog

License

MIT.