Agent RuntimePortable execution semantics for AI agents.

Submit work, stream facts, control tools, resume state, route models, manage tasks, and export evidence without binding to one UI or provider.

Read the specification

Implementation quickstart

Standards ecosystem

LLM full context

Execution authority

Sessions, threads, turns, tasks, runs, steps, tools, actions, subagents, artifacts, and evidence are represented as typed runtime facts.

Control plane included

Submit, interrupt, resume, queue, steer, action response, task operations, subagent control, snapshots, and exports are standard boundaries.

Task-native runtime

Objectives, attempts, dependencies, progress, ownership, delivery state, and peer task refs are first-class facts.

Provider neutral

OpenAI, Anthropic, local models, graph runtimes, terminal agents, and custom streams can adapt into the same event model.

Durable recovery

Snapshots, read models, pending requests, incidents, queues, history windows, and replay exports survive restarts and migrations.

Trust-ready telemetry

Trace ids, spans, cost, limits, evidence refs, replay cases, and audit exports share stable correlation ids.

What Agent Runtime Defines

Contract	What it answers
Runtime event stream	What happened, in which order, and under which ids?
Control plane	Which writes are allowed, and who owns submit, interrupt, resume, queue, task, and action changes?
Agent task	What work is planned, running, blocked, retried, delegated, accepted, or delivered?
Tool and context	Which tools, capabilities, context refs, policies, and permissions were available?
State snapshots	What can be hydrated, inspected, resumed, or repaired later?
Evidence and replay	Which execution facts can be exported for review, replay, audit, and support?

Start Here

For AI Clients

llms.txt: concise navigation index.
llms-full.txt: current English core documentation in one file.
llm.txt and llm-full.txt: compatibility aliases.

Agent Standards Ecosystem

Runtime owns execution facts and controls. UI projects them, Knowledge feeds context, Evidence packages trust records, and adjacent tool/artifact/policy systems keep their own authority.