atif

Pydantic models for the Agent Trajectory Interchange Format (ATIF), a standardized JSON schema for logging the complete interaction history of autonomous LLM agents (user messages, agent responses, tool calls, observations, metrics, and embedded subagent trajectories). Implements ATIF v1.8. This project is independent and not an official implemention.

Install

pip install atif

Usage

from atif import Trajectory

data = {
    "schema_version": "ATIF-v1.8",
    "agent": {"name": "my-agent", "version": "1.0.0", "model_name": "claude-sonnet-4-6"},
    "steps": [
        {"step_id": 1, "source": "user", "message": "What time is it?"},
        {
            "step_id": 2,
            "source": "agent",
            "message": "Let me check.",
            "tool_calls": [
                {"tool_call_id": "c1", "function_name": "now", "arguments": {}}
            ],
            "observation": {
                "results": [{"source_call_id": "c1", "content": "2026-05-16T12:00:00Z"}]
            },
            "metrics": {"prompt_tokens": 120, "completion_tokens": 18},
        },
    ],
}

trajectory = Trajectory.model_validate(data)
print(trajectory.steps[1].tool_calls[0].function_name)  # "now"
print(trajectory.to_json_dict(exclude_none=True))

Validation enforces the ATIF rules: sequential step_ids, ISO 8601 timestamps, agent-only fields gated on source == "agent", tool-call, observation correlation, embedded-subagent trajectory_id uniqueness, and ContentPart text/source XOR with the source kind matching type.

Versioning

atif's MAJOR.MINOR tracks the ATIF RFC version it implements (1.8.x ⇔ ATIF v1.8); PATCH is reserved for library-only fixes.

AI disclosure

Parts of this project, including code, tests, and documentation, were written with the assistance of AI tooling. Readers should be aware that AI was involved in its production.