Threshold: open-source agent harness keeps the project alive while sessions come and go
key_of_door · reddit · 2026-10-10
After four months with coding agents (mostly Codex), the author found himself avoiding fresh sessions on long projects because each session's accumulated context became an asset. His answer is Threshold, an open-source, project-rooted agent harness built on Pi, structured as Project → Task → Run: each Run is a disposable agent that inspects the project, reads its task, checks Git, and resumes from checkpoints and messages left by earlier Runs.
Key design lessons: an early "record everything" approach got too complicated, so he refactored around only giving agents re-verifiable facts (files, Git state, tests) — an agent's account doesn't automatically become truth. Fresh Runs sometimes even catch mistakes in a previous Run's checkpoint by re-inspecting the project. Multiple Runs coordinate via messages and a shared board, schedulers and tool sets are per-Run configurable and replaceable.
Still alpha; he's looking for feedback on where fresh Runs struggle to pick up work.
More from coding & agent
- Index Network open-sources Social Discovery Protocol: agents negotiate fit on your behalf — carsonfarmer · 2026-10-23
- Installed a formal-verification skill: it silently installed Java, MacBook runs 10 degrees hotter — zealcaiden · 2026-10-23
- OpenAI's Decisions API hits GA: near real-time routing up to 10x faster than GPT-6 Luna — OpenAIDevs · 2026-10-23
- Ruff author warns: upgrade Node/Bun now — --use-system-ca still accepts the eMudhra cert — charliermarsh · 2026-10-23
- Proof of Agent: paste one prompt and your AI mines, grades, and self-governs — adamamcbride · 2026-10-23
- DeepLearning.AI weekly: GPT-6.1 hits 97.9% byte parity on COBOL migration — DeepLearningAI · 2026-10-23