🐴 Bridle

Harness an agent into a deterministic program with typed I/O.

Bridle is a Python library for writing agents the way you'd write any program: control flow in code, judgment in the model. Three primitives are typed holes the model fills with values — step, branch, loop. Two decorators compose them: @agent, @tool. Wrappers add behavior: cache, retry, timeout, with_model, fallback, mock, log.

It reads like async/await for LLM decisions.

Status

v0.1.0 — Anthropic-only, sync, single-agent. See What's next for the v0.2.0 roadmap.

Install

pip install bridle-ai

The PyPI distribution is bridle-ai; the import is bridle.

Quickstart

import bridle
from bridle import agent, branch, cache, loop, retry, step, tool
from bridle.models.anthropic import install
from pydantic import BaseModel


class Topic(BaseModel): title: str
class Plan(BaseModel): topics: list[Topic]
class Source(BaseModel): url: str; summary: str
class Brief(BaseModel): headline: str; body: str


@tool
def search(query: str) -> list[str]:
    """Search the web. Returns up to 10 result URLs."""
    ...


@agent(input=str, output=Brief, model="claude-sonnet-4-6")
def brief_writer(topic: str) -> Brief:
    plan = cache(step("draft a research plan", schema=Plan, context=topic))

    sources: list[Source] = []
    for t in plan.topics:
        found = loop(
            f"gather sources on {t.title}",
            schema=Source,
            until=lambda acc: len(acc) >= 3,
            tools=[search],
        )
        sources.extend(found)

    if not branch("is the evidence sufficient?", context=sources):
        return brief_writer(f"{topic} — go deeper on whatever's underdocumented")

    return retry(step("write the brief", schema=Brief, context=(topic, sources)), attempts=2)


install()  # registers the Anthropic adapter as the active model client
result = bridle.resolve(brief_writer("the weather on Mars"))
print(result.headline)

The model never picks the next state. It produces typed values; your Python decides where to go next. Every primitive is mockable; every run is observable.

The four primitives

@tool registers a Python function as a tool the model can call. The parameter schema is extracted from type hints; the docstring becomes the description.

The wrapper algebra

Every wrapper takes a Call and returns a Call. They compose freely.

cache(retry(timeout(step("..."), seconds=10), attempts=3))

Model selection

Bridle ships zero default models. Set one of three layers:

bridle.configure(model="claude-sonnet-4-6")     # process-wide
@agent(model="claude-opus-4-7", ...)             # per-agent
with_model(step("..."), "claude-haiku-4-5")     # per-call (highest precedence)

Resolution order: per-call → per-agent → process. If none is set, Bridle raises ConfigurationError with a message that lists all three places.

The trace

Every primitive emits structured events into a Trace you can inspect, replay, or stream.

from bridle import Trace
from bridle.trace import set_active_trace

trace = Trace()
set_active_trace(trace)

bridle.resolve(brief_writer("..."))

print(trace.to_jsonl())          # one JSON line per event
print(trace.tree())              # nested view: agent → step → model_request → ...
trace.subscribe(lambda e: ...)   # live observer

Event kinds: call_start, call_end, model_request, model_response, tool_call, tool_result, cache_hit, cache_miss, retry.

Token budgets

Soft per-agent budget — usage accumulates after each model response and the next step raises TokenBudgetExceededError if it would breach.

@agent(input=Q, output=R, model="claude-sonnet-4-6", token_budget=100_000)
def deep_research(q: Q) -> R:
    ...

Caching

Caching is opt-in per call via the cache wrapper. Backends ship for memory and file; Redis is reserved for v0.2.0.

from bridle.cache.file import FileCache
bridle.set_cache(FileCache("./.bridle-cache"))

# Now any cache(...) wrapped call writes to disk.
plan = cache(step("draft a plan", schema=Plan, context=topic))

The default key is deterministic across runs: it hashes the call's kind, schema fingerprint, context, prompt, and tools.

What's next

v0.2.0 (planned):

• Async-first execution
• parallel primitive (returning to the four primitives the brief originally proposed, native to async)
• Multi-agent coordination
• Streaming primitives
• Redis-backed cache
• Sealed inner traces (seal=True on @agent)
• Model abstraction beyond Anthropic
• True durable execution for long-running agents

License

MIT.