AgentLoom

Build and run multi-agent applications from YAML.

Typed Workers, permissioned edits, resumable Runs, explicit Goal completion, and review-gated memory share one runtime truth.

AgentLoom treats a multi-agent system as an Application with an execution contract. YAML defines the Supervisor, typed Workers, models, tools, Skills, Hooks, permissions, and runtime policy. Run Applications through the CLI or Python API and inspect structured execution evidence.

Why AgentLoom

Workers become typed tools

A Supervisor explicitly selects Workers through worker_agents; each selected Worker becomes a callable tool named and described by that Worker. Simple Workers use an empty Tool argument object and text output; each Worker's task comes from its YAML. Complex Workers declare Draft 2020-12 input_schema and output_schema, which runtimes validate as executable contracts rather than prompt conventions.

Runs produce evidence, not terminal guesses

Every allocated Run receives an immutable run_id, manifest, and versioned lifecycle events, with bounded file logs when enabled plus audit records and artifacts. A logical task_id survives resume. The CLI JSON/JSONL and Python API read the same canonical state. Preflight rejection occurs before a Run or its storage is allocated.

Long-running work has an explicit owner

Goal Mode keeps one root Supervisor objective active across continuation segments and Worker delegation. Only that Supervisor can mark the Goal complete with evidence. Checkpointing preserves the Goal and completed work across interruptions and resume.

Memory has review boundaries

Self-Learning v6 stores searchable history and evidence-gated memory separately. Fact and experience candidates pass evidence gates and the configured scope-approval policy; promotion to Project scope is always initiated by a person.

Extensions do not silently gain authority

Skills are model-context packages loaded on demand. Hooks are separately and explicitly authorized runtime code. Built-in tool metadata is discoverable without importing implementations, while actual tool, file, Shell, and MCP access remains governed by Agent configuration and permissions.

Quick start

The source installer uses uv to create an isolated Python environment with locked dependencies and installs the agentloom and loom commands for this checkout:

git clone https://github.com/linora-u/AgentLoom.git
cd AgentLoom
./install

It currently supports macOS and Linux shells and requires Git and Bash. It installs missing uv through its official installer, then places the environment in this checkout's .venv and the launchers/install records in .agentloom. The project pins its managed Python version in .python-version; ./install uses that version even when another Python is available on PATH. No environment activation is needed. Set AGENTLOOM_INSTALL_DIR=/path/to/install for a custom installation with its environment in /path/to/install/venv. Open a new terminal and verify it:

agentloom --help
loom --help

The source installer defaults to smol plus professional code tools. For Pi without smol, use ./install --runtime pi with Node 22.19+ and npm available; it automatically downloads the pinned SDK and builds AgentLoom's bridge. See runtime installation profiles for locked checkout/release commands and clean Application verification.

Create the local model configuration:

cp config/model.example.yaml config/model.yaml

model:
  default_model_type: powerful
  powerful:
    model: "openai/<model-id>"
    api_key: "<api-key>"
    base_url: "https://<openai-compatible-endpoint>"  # optional for OpenAI
    tool_choice: "auto"
  fast:
    model: "openai/<fast-model-id>"
    api_key: "<api-key>"
    base_url: "https://<openai-compatible-endpoint>"
    tool_choice: "auto"

config/model.yaml is ignored by Git and provides the project's model catalog. Applications can also define their own catalog in config/model.yaml under the Application directory. Select models in the Agent YAML with model_type.

Define an Application

An Application keeps its Supervisor, Workers, prompts, optional tools, and outputs together:

applications/release_review/
├── workflows/
│   ├── release_review_agent.yaml
│   └── worker_agents/
│       ├── api_reviewer.yaml
│       └── test_reviewer.yaml
├── config/system.yaml          # optional Application overlay
├── skills/                     # optional private Skills
└── sysprompt/                  # optional prompt templates

Backend-specific execution settings belong only in runtime_options. Historical top-level smol fields are silently ignored without conversion or rejection.

A Supervisor references Worker definitions:

name: "release_review"
agent_runtime: "smolagents"
description: "Review an API release and its test evidence."
model_type: "powerful"

worker_agents:
  - path: "workflows/worker_agents/api_reviewer.yaml"
  - path: "workflows/worker_agents/test_reviewer.yaml"

task: |
  Ask both Workers for evidence, reconcile conflicts, and return one release decision.

tools: []
runtime_options:
  max_steps: 12
goal:
  enabled: true

Each Worker has its own YAML task. Without input_schema, its Tool takes no arguments. Declare schemas when the Supervisor passes typed call data or needs structured output:

name: "api_reviewer"
agent_runtime: "smolagents"
description: "Review API compatibility risks."
model_type: "fast"

input_schema:
  type: object
  properties:
    request:
      type: string
      description: "Release scope and API diff."
  required: [request]
  additionalProperties: false

output_schema:
  type: object
  properties:
    decision:
      type: string
      enum: [compatible, incompatible]
    findings:
      type: array
      items:
        type: string
  required: [decision, findings]
  additionalProperties: false

task: |
  Review the request, cite evidence, and return prioritized findings.

tools: []
worker_agents: []
runtime_options:
  max_steps: 8

Run the Supervisor directly:

uv run --locked --extra smol --extra code loom run applications/release_review/workflows/release_review_agent.yaml

Or ask a Skill-aware coding assistant to read agentloom-framework-skill/SKILL.md, create the files, validate them, run the Application, and inspect .agentloom evidence.

Persistent Worker queues

The optional durable extra uses DBOS 3.2.0 for persistent queues, admission, retries, and recovery. Its default storage is a local SQLite file. Declare each Worker's shared queue capacity in the Application's config/system.yaml:

worker_execution:
  workers:
    workflows/research.yaml:
      concurrency: 4

Submit through an Application Run, using a stable namespace root and an explicit business revision:

from agentloom.app import Application
from agentloom.execution.concurrency import WorkItem

with Application.open(application, project_root=project_root) as app:
    with app.durable(root=namespace_root) as execution:
        handles = execution.worker("workflows/research.yaml").submit_many(
            [WorkItem("research-1", {"query": "Investigate the release."})],
            revision="1",
        )

Start one dedicated consumer for that namespace. Other processes can submit to the same queues; they share the declared capacities. Finite worker.map() calls retain their separate per-call limits.

uv run --locked --extra durable loom durable serve APPLICATION \
  --worker workflows/research.yaml --root NAMESPACE_ROOT

An existing task with different immutable input is rejected. Each invocation runs a complete Application under macOS launchd, and resource exit is checked before replacement. Generic execution requires explicit host authorization to replay work; applications that commit business side effects can supply a trusted host with completion receipts. The current OS adapter supports macOS; Linux and actual PostgreSQL or multi-host execution have not been validated. See the execution contract.

Runtime model

The Python runtime owns model routing, Worker-tool generation, concurrency, permissions, Hooks, checkpoints, and evidence. Deterministic preprocessing, validation, caching, and output writing remain ordinary Python code.

Runtime storage separates attempts from recoverable tasks:

.agentloom/
├── runs/<application_id>/<run_id>/
│   ├── manifest.json
│   ├── logs/runtime.log
│   ├── audit/
│   └── artifacts/
├── checkpoints/<application_id>/<task_id>/
│   ├── checkpoint.json
│   ├── workers/<worker>/calls/<index>/checkpoint.json
│   ├── todos.json
│   ├── goal.json
│   ├── context_store/
│   └── file-history/
└── workspaces/agents/<application_id>/<agent_path>/
    ├── insights.md
    └── tasks/<task_id>/{context.md,trace.md}

Goal, Todo, context-store, file-history, and Recall files appear only when the corresponding feature is configured or used.

Run and integrate

Run the included code-review Application without creating a new Application:

uv run --locked --extra smol --extra code loom run applications/ai_quality_analysis/workflows/code_review_agent.yaml

Use machine-readable lifecycle events when another program owns execution:

uv run --locked --extra smol --extra code loom run applications/ai_quality_analysis/workflows/code_review_agent.yaml --output-format json
uv run --locked --extra smol --extra code loom run applications/ai_quality_analysis/workflows/code_review_agent.yaml --output-format jsonl

For programmatic execution, execute_app() returns an ApplicationRunResult with output, timestamps, structured Goal state, and a RunInfo receipt:

from agentloom.app.runner import execute_app

result = execute_app("applications/release_review/workflows/release_review_agent.yaml")
print(result.output, result.run.run_id)

Framework source lives directly in src/app/, src/execution/, src/integrations/, and the other responsibility modules. Installation maps src/ to the Python package name agentloom: the import above loads src/app/runner.py. There is no extra agentloom source directory. Use uv sync --managed-python --locked --all-groups --extra smol --extra code to install the checkout before calling Python APIs. The old src.* imports and module commands are removed; supported commands are agentloom, loom, and python -m agentloom. See the architecture and migration map for module ownership and migration details.

Post-allocation failures carry the same receipt; preflight rejection emits run.rejected before storage exists. See Structured Run API.

For finite parallel work, use app.worker(path).map(items). For persistent tasks, install the durable extra and submit through the public DBOS adapter:

from agentloom.app import Application

with Application.open("applications/release_review", project_root=".") as app:
    with app.durable(root=".agentloom/release-review-queue") as execution:
        handles = execution.worker("workflows/research.yaml").submit_many(items, revision="1")

Consume the same namespace with loom durable serve applications/release_review --worker workflows/research.yaml --root .agentloom/release-review-queue. DBOS 3.2.0 owns queues and recovery; SQLite is the default. Each invocation runs under macOS launchd. Linux durable execution and remote database operation have not been validated. See the execution specification and full-day acceptance.

Durable schedules use the same Application contract and Run lifecycle. Their automatic firing is handled by an explicit foreground service:

agentloom schedules --project /path/to/project serve

Example Applications

Application	Demonstrates
ai_quality_analysis	Twelve specialized Workers coordinated into staged code review
repo_map	Deterministic preprocessing, bottom-up Agent analysis, batching, and progress persistence
goal_mode_validation	Explicit Goal completion, continuation, and checkpoint resume
self_learning_smoke	Session history, memory proposals, evidence, and review boundaries

Documentation

Document	Covers
Configuration Overview	Configuration layers, merging, and isolation
Agent Configuration	Supervisor and Worker YAML fields
Tool Catalog	Lazy implementation loading, toolsets, metadata, and extension rules
Skills	Discovery, on-demand activation, and permission boundaries
Hooks	Explicit authorization, events, transforms, and failure semantics
Goal Mode	Continuation, completion ownership, resume, and schedules
Checkpoint and Runtime Storage	Run/task identity, evidence, recovery, and retention
Self-Learning v6	History, candidates, review, approval, and promotion
Structured Run API	Python receipts, typed failures, JSON, and JSONL

Development and support

# Framework
uv run --locked --extra smol --extra code pytest tests -q

• Issues: github.com/linora-u/AgentLoom/issues
• Contact: [email protected]

If AgentLoom helps your project, consider starring the repository or contributing a focused Application, fix, or validation case.