LRA — Long-Running Agents

A durable, self-improving agent organization that makes verified progress on software missions spanning days to weeks — surviving crashes, reboots, and context-window limits without losing the plot.

Durable execution (Temporal) · git as source of truth · deterministic verification · pluggable models · runs at $0

---

**Status: early / work-in-progress (Phase 0).** The durable core is being built first. See [the roadmap](#roadmap) and [what's proven vs. frontier](#honesty-what-is-proven-vs-frontier). This README describes the system as designed; features are landed incrementally and nothing here is marketing — see the honesty section.

What this is

Most "autonomous agents" are a single LLM loop that dies the moment the process restarts, the context window fills, or an API call fails. **LRA treats long-horizon autonomy as an engineering problem, not a model capability.** The model thinks in short bursts; the *system* runs for weeks by:

• Externalizing the source of truth. The context window is a lossy cache. The real state — a git repo + a structured progress/checklist/decision log — lives outside it and is re-read every cycle ("assume interruption"). A reboot on day 12 reconstructs situational awareness in seconds.
• Durable execution. The agent loop runs inside a Temporal workflow. Every LLM/tool call is a journaled, retried, replay-from-cache activity, so a crash resumes exactly where it left off — no work lost, no tokens re-spent. Idle time is spent in durable sleep at zero cost.
• Deterministic verification. A checklist item is only "done" when real tests/lint/build/ typecheck pass — never on the model's say-so. This is what stops per-step errors compounding over thousands of steps.