Run Claude agents for hours — one agent, or a whole fleet — without any of them silently dying.

Yes. Apache-2.0, open source, hosted on GitHub. There is no paid tier and no licence to negotiate. AI Heroes makes money when teams want it set up and integrated for them, not from the code itself.

Today, yes. The harness is built around Claude Code's hooks system, the Read tool, and the turn-boundary events. The two-pulse pattern is portable in principle, but this implementation depends on Claude Code primitives.

Two clocks. The inner pulse runs at every turn boundary inside the agent and checks whether the work is done. The outer pulse runs every 15 minutes on a real wall clock and checks whether the inner pulse is still alive. If the inner pulse stops writing heartbeats, the outer pulse alerts and writes recovery guidance the agent will read next.

Yes. Beyond single-agent mode, one Claude Opus orchestrator can drive a fleet of Claude and Codex subagents working in parallel — each in its own isolated git worktree so they never collide. The same 15-minute heartbeat, the same objective evaluator agent, and the same kill switch watch every agent in the fleet at once. You run one long goal with one agent, or many lanes with a whole team, on the same control system.

A Claude Opus model runs the show — it owns the planning, the review, and the accountability for whether the goal is met. It delegates the heavy execution to Codex agents, which are fast at writing and changing code. Opus decides what good looks like and checks the result; Codex does the volume. You get senior judgement on top of high-throughput execution, instead of one model trying to do both jobs at once.

Every success criterion starts as false. The only way to flip a criterion to true is fresh evidence read through Claude Code's Read tool and validated by the verify-gate hook. This stops the agent from declaring a goal complete without showing its work. It is the single most important safeguard in the harness.

Goals with a real programmatic terminator. Engineering work with a test suite, migrations gated on build and lint, content batches with an automated audit, multi-sprint product work with a measurable success gate. If you cannot write a check that returns true or false without a human, the harness has nothing to wait for.

Yes. It extends Anthropic's primitives directly — the Effective harnesses for long-running agents post (November 2025), the Harness design for long-running application development post (March 2026), and the anthropics/cwc-long-running-agents reference repository. AI Heroes adds the outer 15-minute stall detector, a pinned Codex executor, and an optional OpenClaw supervisor on top.

Yes. The harness is free, but installing it cleanly, wiring it into your Claude Code launchers, choosing the right goals, and running a first end-to-end overnight session is work most teams would rather not do themselves. Book a call from this page and we will scope it.

Drop an AGENT_STOP file into the workspace. The harness stops at the next turn boundary, cleanly. No process killing, no orphaned state. Every other operator control — STEER.md for mid-run steering, the heartbeat audit trail, the session ledger — is documented in the repository README.