agentloop

Python SDK for AgentLoop — middleware that turns human corrections into searchable memory for your agent.

• Python 3.9+, sync and async
• One runtime dependency (httpx), nothing else
• Graceful by default — network blips don't break your agent
• Type hints throughout, full dataclass response shapes

Install

pip install agentloop-py

Quick start (sync)

import os
from openai import OpenAI
from agentloop import AgentLoop

loop = AgentLoop(api_key=os.environ["AGENTLOOP_API_KEY"])
openai = OpenAI()

def ask(question: str, user_id: str) -> str:
    # 1. Before calling the LLM, pull relevant corrections.
    memories = loop.search(question, user_id=user_id, limit=3)

    # 2. Inject them into your system prompt.
    facts = "\n".join(f"- {m.fact}" for m in memories) if memories else ""
    system = "You are a helpful assistant."
    if facts:
        system += f"\n\nTrusted facts from past corrections:\n{facts}"

    # 3. Call the LLM.
    resp = openai.chat.completions.create(
        model="gpt-4o-mini",
        messages=[
            {"role": "system", "content": system},
            {"role": "user", "content": question},
        ],
    )
    answer = resp.choices[0].message.content or ""

    # 4. Log the turn for review.
    loop.log_turn(question, answer, user_id=user_id)

    return answer

That's the whole integration — one call before the LLM, one after. For a zero-setup drop-in wrapper, install agentloop-py-openai or agentloop-py-anthropic instead.

Quick start (async)

from agentloop.aio import AsyncAgentLoop

async def ask(question: str, user_id: str) -> str:
    async with AsyncAgentLoop(api_key=...) as loop:
        memories = await loop.search(question, user_id=user_id)
        # ... LLM call ...
        await loop.log_turn(question, answer, user_id=user_id)

Use AsyncAgentLoop from agentloop.aio in asyncio code (FastAPI, aiohttp, etc.). Same API, all methods are coroutines except feedback_url() which stays sync — no network call, just HMAC.

Configuration

loop = AgentLoop(
    api_key="ak_...",                           # required
    base_url="https://...",                     # default: https://api.getagentloop.io
    timeout_s=10.0,                             # per-request timeout
    feedback_signing_secret="...",              # only needed for feedback_url()
    throw_on_error=False,                       # see "Graceful failures"
    http_client=my_httpx_client,                # inject custom httpx.Client
)

Base URL resolution

The base_url argument follows this priority order:

1. Explicit base_url passed to the constructor (highest)
2. AGENTLOOP_BASE_URL environment variable
3. Hardcoded hosted-backend URL (fallback)

This lets you point the SDK at a local dev server, a self-hosted deployment, or a future gateway without code changes — just set AGENTLOOP_BASE_URL in the environment.

export AGENTLOOP_BASE_URL=http://localhost:8080
python your_app.py  # now points at a local dev backend instead of api.getagentloop.io

Four methods

search(query, *, user_id=None, limit=3, tags=None) → list[Memory]

Retrieve corrections before your LLM call. Returns [] on failure (unless throw_on_error=True).

memories = loop.search("pix limit at night", user_id="u_123", limit=5)
for m in memories:
    print(m.fact, m.score, m.tags)

log_turn(question, agent_response, *, user_id=None, session_id=None, signals=None, metadata=None) → LogTurnResponse

Queue a turn for review. Returns a default response on failure (unless throw_on_error=True).

result = loop.log_turn(
    question="What's the Pix limit?",
    agent_response="R$5,000",
    user_id="u_123",
    signals={"thumbs_down": True},
    metadata={"latency_ms": 240, "model": "gpt-4o-mini"},
)
if result.was_duplicate:
    # Backend deduplicated this turn against an existing pending one.
    # The returned turn_id points to the merged doc.
    ...

annotate(*, question, agent_response, correction, rating, ...) → AnnotateResponse

Create an annotation directly (bypass the review queue). Always throws on failure — silent degradation would hide reviewer work.

result = loop.annotate(
    question="What's the Pix limit at night?",
    agent_response="R$5,000",
    correction="Pix limit between 8pm and 6am is R$1,000.",
    rating="incorrect",
    root_cause="context",
    tags=["pix", "limits"],
    reviewer="maria@luma.com.br",
)

feedback_url(question, agent_response, *, user_id="", session_id="") → str

Generate an HMAC-signed URL for the embedded feedback widget. Requires feedback_signing_secret. Sync on both clients.

url = loop.feedback_url(question, answer, user_id="u_123")

Signatures are byte-identical to the JavaScript SDK, so URLs signed in Python validate on backends that also accept JS-signed URLs.

Graceful failures

By default, search() and log_turn() return empty/default values on failure. Agent-loop calls sit on the critical path of your agent's response; a blip shouldn't turn into a 500.

# Hard failures
from agentloop import AgentLoopError

loop = AgentLoop(api_key=..., throw_on_error=True)
try:
    memories = loop.search("...")
except AgentLoopError as e:
    print(e.status, e.body)

annotate() always raises regardless of this flag.

Context managers

Both clients support context management for clean resource cleanup:

with AgentLoop(api_key=...) as loop:
    loop.search("...")
# httpx.Client is closed on exit

async with AsyncAgentLoop(api_key=...) as loop:
    await loop.search("...")
# httpx.AsyncClient is closed on exit

License

MIT