AgentLane

AgentLane is a runtime-first framework for building reliable, inspectable AI agent systems.

AgentLane is for AI workflows where agent behavior is part of the application architecture. It gives agents stable identities, routes work through explicit messages, and lets a local agent loop grow into background workers, pub/sub flows, and distributed runtimes without changing the core communication model.

Most agent frameworks start with a prompt, a few tools, and a loop. AgentLane starts one layer lower: runtime, addressed messaging, delivery outcomes, and agent instance reuse. Model calls, tools, handoffs, and the default harness sit on top of that runtime foundation.

That shape matters when users depend on the system. You need to know which agent handled work, where state lives, which messages and tools were involved, how work was delegated, and how the workflow can be tested, reproduced, and operated.

What You Get

AgentLane is organized into layers that can be used together or independently:

1. Runtime and Messaging — addressed agents, direct sends, scheduling, pub/sub, delivery outcomes, local execution, and distributed workers.
2. Models — prompt templates, schemas, structured outputs, native tools, and provider clients.
3. Harness — DefaultAgent, markdown agent definitions, resumable run state, tool execution, handoffs, agent-as-tool delegation, shims, skills, and a local stdio process bridge for TypeScript app shells.
4. Transport — wire-safe serialization boundaries for distributed payloads.
5. Tracing — observability across runtime, model, and harness execution.

These layers let you start with a simple local agent and keep the same runtime model as the workflow grows into addressed services, background specialists, fan-out and fan-in, or distributed execution.

When To Use AgentLane

Use AgentLane when you are building AI systems that need one or more of:

1. local agents with tools, handoffs, delegation, or resumable runs
2. stable identities for agents, services, and background specialists
3. explicit routing between model-backed agents and deterministic workers
4. fan-out, fan-in, pub/sub, or human-review workflows
5. structured model calls with schemas, tools, and provider adapters
6. a path from local development to distributed execution
7. orchestration that stays in application code

AgentLane is especially useful when the agent workflow is part of the product architecture and carries responsibilities beyond a single model call.

╔════════════════════════════════════════════════════════════════════════════════════╗
║                                                                                    ║
║    █████╗  ██████╗ ███████╗███╗   ██╗████████╗██╗      █████╗ ███╗   ██╗███████╗   ║
║   ██╔══██╗██╔════╝ ██╔════╝████╗  ██║╚══██╔══╝██║     ██╔══██╗████╗  ██║██╔════╝   ║
║   ███████║██║  ███╗█████╗  ██╔██╗ ██║   ██║   ██║     ███████║██╔██╗ ██║█████╗     ║
║   ██╔══██║██║   ██║██╔══╝  ██║╚██╗██║   ██║   ██║     ██╔══██║██║╚██╗██║██╔══╝     ║
║   ██║  ██║╚██████╔╝███████╗██║ ╚████║   ██║   ███████╗██║  ██║██║ ╚████║███████╗   ║
║   ╚═╝  ╚═╝ ╚═════╝ ╚══════╝╚═╝  ╚═══╝   ╚═╝   ╚══════╝╚═╝  ╚═╝╚═╝  ╚═══╝╚══════╝   ║
║                                                                                    ║
║                reliable, inspectable AI agent workflows                            ║
║                                                                                    ║
║              runtime • messaging • model primitives • harness                      ║
║                                                                                    ║
║                 from local agents → distributed agent systems                      ║
║                                                                                    ║
╚════════════════════════════════════════════════════════════════════════════════════╝

Installation

Install AgentLane with uv:

uv add agentlane

If you are trying the repository directly instead:

uv sync --all-extras

Quick Start

Define an agent in a markdown file — frontmatter for config, the body as its system prompt — and run it. The four steps below are one progression: the same agent grows from a single file to a coordinated team, and from one local process to a distributed runtime, without changing the model or the run loop.

All four reuse one model client:

import asyncio
import os

from agentlane_openai import ResponsesClient

from agentlane.harness.agents import DefaultAgent
from agentlane.models import Config

model = ResponsesClient(
    config=Config(api_key=os.environ["OPENAI_API_KEY"], model="gpt-5.4-mini"),
)

1. An agent from a markdown file

care_navigator.md:

---
name: care-navigator
description: Guides patients to a clear next step for a new symptom or concern.
---
You are a concise patient care navigation agent. Give one clear next step. When
a clinical tool is available, use it before advising on a medication.

async def main() -> None:
    agent = DefaultAgent.from_markdown("care_navigator.md", model=model)
    result = await agent.run(
        "I feel dizzy after starting a new blood-pressure medication. What first?"
    )
    print(result.final_output)


asyncio.run(main())

One file, one run(...). The frontmatter configures the agent; the body is its system prompt. It runs on a local single-threaded runtime by default, and each run(...) stores resumable state on the agent.

2. Two markdown agents on a distributed runtime

Add a specialist as a sub-agent and bind the pair to a distributed runtime. The specialist becomes an addressed agent the lead delegates to — ready to move onto its own worker later. Only subagents= and runtime= change.

med_safety.md:

---
name: med-safety
description: Use to check a medication for interactions and safety flags before advising.
model: inherit
---
You review a medication for interactions and safety flags, and return a short
note that says clearly when something is urgent.

from agentlane.runtime import distributed_runtime


async def main() -> None:
    async with distributed_runtime() as runtime:
        agent = DefaultAgent.from_markdown(
            "care_navigator.md",
            model=model,
            subagents=["med_safety.md"],
            runtime=runtime,
        )
        result = await agent.run(
            "I started lisinopril yesterday and feel lightheaded. Is that expected?"
        )
        print(result.final_output)


asyncio.run(main())

The lead calls the med_safety specialist as a tool, gets its result, and answers. model: inherit lets the specialist reuse the lead's model.

3. Full programmatic control

When you need real Python tools, tuned model calls, and run-loop limits, build the descriptor directly. A plain function becomes a tool — no decorators or registration:

from agentlane.harness import AgentDescriptor, Runner
from agentlane.models import Tools


def lookup_medication(name: str) -> str:
    """Return basic guidance for a medication by name."""
    return f"{name}: take with food; report severe dizziness or fainting to your care team."


async def main() -> None:
    agent = DefaultAgent(
        descriptor=AgentDescriptor(
            name="care-navigator",
            model=model,
            instructions="Call lookup_medication before advising on a medication. Give one next step.",
            tools=Tools(tools=[lookup_medication]),
            model_args={"reasoning_effort": "low"},
        ),
        runner=Runner(max_turns=6),
    )
    result = await agent.run("Can I keep taking metformin if it upsets my stomach?")
    print(result.final_output)


asyncio.run(main())

The markdown agent and this one share the same DefaultAgent contract — markdown is the fast path, the descriptor is the full surface.

4. A multi-agent team on a distributed runtime

Attach specialists with subagents= and run the team on a distributed runtime. DefaultAgent wires each sub-agent in for you — no manual as_tool(). The coordinator owns the model loop; each specialist runs as an addressed agent the runtime can place on its own worker:

med_safety = AgentDescriptor(
    name="med-safety",
    model=model,
    tools=None,
    instructions="Flag medication interactions or safety concerns in one sentence.",
)
guidelines = AgentDescriptor(
    name="guidelines",
    model=model,
    tools=None,
    instructions="Cite the relevant care guideline for the symptom in one sentence.",
)


async def main() -> None:
    async with distributed_runtime() as runtime:
        triage = DefaultAgent(
            descriptor=AgentDescriptor(
                name="triage-lead",
                model=model,
                instructions="Send medication questions to `med_safety` and symptom questions to `guidelines`, then give one next step.",
            ),
            subagents=[med_safety, guidelines],
            runtime=runtime,
        )
        result = await triage.run(
            "New chest tightness after a dose increase of my heart medication. What should I do?"
        )
        print(result.final_output)


asyncio.run(main())

subagents= is the one way to attach sub-agents — markdown paths via from_markdown (step 2) or descriptors here — and tools=[child.as_tool()] stays as a manual escape hatch. The same runtime concepts scale on to background specialists, pub/sub fan-out, and multi-process workers — see Harness Distributed Agents.

Repository examples

If you are running from a repository checkout, run one runtime example:

uv run python examples/runtime/multi_agent_workflow/main.py

Run one high-level harness example with a real model:

OPENAI_API_KEY=sk-... uv run python examples/harness/default_agent_quickstart/main.py

Run the distributed harness agent smoke test:

uv run python examples/harness/distributed_clinical_inbox_copilot/main.py \
  --multiprocess \
  --smoke-review

The runtime example shows explicit message passing. The distributed harness example shows a top-level agent coordinating worker runtimes through publish-based fan-out and fan-in.

Choose the layer you need

Runtime

Use the runtime when agent identity, message routing, pub/sub, scheduling, or distributed execution are part of your application design.

Start here:

1. Runtime: Engine and Execution
2. Runtime: Distributed Runtime Usage
3. Messaging: Routing and Delivery

Models

Use the models layer when you want reusable prompt templates, schemas, structured outputs, tools, or provider clients without adopting the full agent harness.

Start here:

1. Overview
2. Prompt Templating

Harness

Use the harness when you want high-level agents, reusable loops, tool execution, handoffs, or agent-as-tool patterns on top of the lower-level primitives.

Start here:

1. Default Agents
2. Markdown Agent Definitions
3. Architecture
4. Tools
5. Shims
6. Compaction
7. Skills
8. Distributed Agents

Documentation

Use the documentation index for the full docs tree:

1. Documentation Index
2. Examples Index
3. Runtime: Distributed Runtime Usage
4. Harness Distributed Agents
5. Tracing Overview
6. Changelog

Origins

AgentLane was initially inspired by Microsoft AutoGen, but takes a runtime-first approach focused on addressed messaging, explicit orchestration, and local-to-distributed execution.

Development

Format, lint, and test:

/usr/bin/make format
/usr/bin/make lint
/usr/bin/make tests

Run one test with:

uv run pytest -s -k <test_name>

Contributing

1. Keep changes small and focused.
2. Add or update tests when behavior changes.
3. Update public docs and examples when the developer-facing surface changes.
4. Ensure formatting, linting, and tests pass before opening a PR.