Contract4Agents

Contract4Agents is a free, open-source toolkit for defining structured agents in external specifications and turning those specifications into usable agent components. It includes a typed contract language, a compiler, code generators, and adapters for OpenAI Agents SDK, Strands Agents, Google ADK, and Amazon Bedrock AgentCore Harness.

Define each agent in its own .agent.contract file. Keep its inputs, output, grants, context, goal, and guidance in that file. Put shared types and capabilities in .types.contract and .capabilities.contract files. Contract4Agents compiles all source files in the project together and derives schemas, documentation, and normal framework objects. Your application supplies the tool implementations and runs the agents.

This gives you one place to read and change the agent design. You can review its guardrails and expectations without finding them across prompts, SDK configuration, and application code. Target bindings select implementations and models without repeating the contract. Validation and evaluation tools help you check that the implementation matches the design.

A Complete Example

This example defines one support agent, connects a Python tool, and runs the agent through the OpenAI Agents SDK.

In a Python 3.11+ project managed by PDM, install the OpenAI adapter:

pdm add "contract4agents[openai]"

Create these files, including an empty your_app/__init__.py:

your_app/
  __init__.py
  tools.py
  run_agent.py
  agent_contracts/
    support.agent.contract
    support.types.contract
    support.capabilities.contract
    contract4agents.targets.toml

1. Define the Agent and Its Tool

Put the shared types in your_app/agent_contracts/support.types.contract:

type KnowledgeResult:
    answer: string
    source_ids: list[string]

type SupportReply:
    answer: string
    source_ids: list[string]
    needs_follow_up: boolean

Put the tool interface in your_app/agent_contracts/support.capabilities.contract:

tool knowledge.search(query: string) -> KnowledgeResult:
    description = "Search the approved support knowledge base."
    side_effect = false

Put the agent definition in your_app/agent_contracts/support.agent.contract:

agent SupportResponder(question: string) -> SupportReply:
    use knowledge.search:
        availability = enabled
        authorization = preapproved
        execution = host

    goal = "Answer the support question from approved evidence."
    description = "Handles first-line support questions."
    guidance = [
        "Use the knowledge tool before answering.",
        "Include the source IDs behind the answer.",
        "Set needs_follow_up when the evidence is insufficient.",
    ]

The tool declaration defines the interface. The use block gives this agent access to it. The agent's return type defines the structure of its answer. Names are global across the project. The agent can reference the shared tool and types directly. File boundaries organize the source; they do not create scope boundaries. Each named file suffix limits which declarations that file can contain. See File Organization for the complete rules.

2. Supply the Implementation and Model

Put the Python tool in your_app/tools.py:

def search_knowledge(query: str) -> dict[str, object]:
    return {
        "answer": "Orders normally leave the warehouse within two business days.",
        "source_ids": ["shipping-policy-2026"],
    }

This example returns fixed data. In your application, replace the function body with a lookup in your knowledge base.

Connect that function and select a model in your_app/agent_contracts/contract4agents.targets.toml:

schema_version = "1"

[targets.openai]
adapter = "openai"

[targets.openai.tools."knowledge.search"]
python = "your_app.tools:search_knowledge"

[targets.openai.profiles.development]
default_model = "gpt-6-luna"

The target binding contains implementation choices. It does not repeat the agent's instructions, tool access, or output schema. Choose a model available to your provider account.

3. Check the Definition

Run these commands from the project root:

pdm run contract4agents check your_app/agent_contracts
pdm run contract4agents compile your_app/agent_contracts --out .contract/build
pdm run contract4agents plan your_app/agent_contracts --target openai --profile development

check validates the source and target bindings. compile writes schemas, instructions, and documentation. plan shows how the selected target will represent the agent and reports missing bindings or unsupported requirements.

These commands do not call a model. Binding checks import Python modules to inspect callable signatures, so module-level code runs during those imports.

4. Run the Agent

Put this in your_app/run_agent.py:

import asyncio

from agents import Runner
from contract4agents import materialize


async def main() -> None:
    system = materialize(
        "your_app/agent_contracts",
        target="openai",
        profile="development",
    )
    agent = system.agents["SupportResponder"]
    run_input = system.serialize_input_for_agent(
        agent,
        {"question": "When will my order ship?"},
    )
    result = await Runner.run(agent, input=run_input)
    print(result.final_output)


if __name__ == "__main__":
    asyncio.run(main())

Set OPENAI_API_KEY in your environment, then run:

pdm run python -m your_app.run_agent

This step makes a provider request. materialize() constructs an ordinary SDK agent with its tool and generated output type. The SDK runs it and returns a structured SupportReply. You write the tool and invocation code; the contract supplies the agent configuration.

Your application continues to own credentials, business rules, approval decisions, sessions, retries, persistence, and deployment.

What Gets Generated?

File generation and runtime construction serve different needs:

materialize() compiles the contract internally; it does not require a prior compile command. It checks the objects it constructs against the target plan. When generated files are committed, compile --check and generate --check can detect stale files in CI.

See Compiler Outputs for file layouts and generation commands.

Framework Support

Install the corresponding extra: contract4agents[openai], contract4agents[strands], contract4agents[google-adk], or contract4agents[agentcore-harness].

The Harness Research example runs with synthetic transport offline. Its live route requires explicit deployment and verification. The application owns approvals, tool execution, and continuation.

Targets have different capabilities. Planning reports those differences and rejects required mappings that would lose declared behavior. For example, Strands and Google ADK do not currently support contract handoffs. The target references describe the supported combinations of tools, approvals, context, and composition. TypeScript generation produces data types and validators; the current native agent adapters run in Python.

Check Behavior Against the Contract

You can add .eval scenarios, capture traces, compare contract changes, and assemble review reports as your application needs them.

• Evals describe scenarios and expected results. The current eval replay command assesses recorded outputs and evidence; it does not run a model.
• Traces connect observed agent and tool activity to the contract and plan.
• Assessment reports whether available evidence supports a requirement: passed, violated, or unverified.
• Semantic diffs and assurance bundles help reviewers inspect design changes and collected results.

These tools support implementation and review. Missing evidence stays unverified. See Capture and Assure a Run for a worked example and Evals, Controls, and Assurance for the assessment rules.

Editor Support

The VS Code extension provides syntax highlighting, completions, diagnostics, hover help, navigation, rename, and optional inlay hints for .contract and .eval files. It uses the package's Python language server, so editor feedback follows the same language rules as the CLI.

Learn More

• First Contract Project: build the support example step by step.
• Using Contract4Agents in an Application: integrate contracts with application code.
• Incident Command: a complete multi-agent example with local fixtures.
• Multi-Lens Research: delegation, typed results, and isolation requirements.
• Market Research Brief: host tools and provider search.
• Documentation Index: tutorials, language and API references, and contributor guides.
• Vision: product purpose and ownership boundaries.

Development

From a checkout of this repository:

Install the native Perfect Doc executable and make it available on PATH. On macOS, use:

brew install btfranklin/tap/perfect-doc

pdm install
npm --prefix tests/typescript ci
npm --prefix editors/vscode ci
pdm run docs-check
pdm run validate
pdm build

pdm run docs-check runs Perfect Doc offline with perfect-doc.toml, then checks repository documentation consistency. pdm run validate includes both documentation checks.

See Validation and Quality Gates for checks by change type and opt-in live tests. Coding agents should start with AGENTS.md.

License

MIT. See LICENSE.