Meta's ReHarness adds backward revision to agent harnesses, beating memory and reflection baselines
Yuxin Xiao · hf · 2026-10-12
Researchers from MIT, Meta MRS Research, and Yale propose ReHarness, a recurrent harness framework for agentic continual learning under a strictly online streaming protocol.
- Existing agent harnesses evolve state only through forward information flow, or need extra replay/validation data. ReHarness complements forward read/write with backward revision: after each instance, the resulting trajectory is used as feedback to revise the reflections retrieved for that instance, propagating feedback backward through the textual harness state.
- Across 3 continual learning benchmarks and 3 language model backbones, ReHarness consistently outperforms strong raw-context baselines, memory- and reflection-based methods, and prompt optimization approaches.
- Analysis shows revisions are source-grounded and meaningful, correct overgeneralized or outdated experience under distribution shifts, and improve both stability and plasticity.
More from coding & agent
- A minimal recipe for AI persona continuity: prefill-injected identity kernel plus a name — RileyRalmuto · 2026-10-12
- What should personal agents do first? Boring renewal and expiry reminders — sujingshen · 2026-10-12
- Rive renders entire 3D city with vector graphics, ships CLI for Codex and Claude Code agents — soleio · 2026-10-12
- TechnoLama shows the AI-recreated PowerPoint lookalike from the ArtCraft suite — technollama · 2026-10-12
- Unpacking agent product forms: capability, collaboration and accountability are separate axes — sujingshen · 2026-10-12
- Dev vibe-codes a minimal Mac app that floats YouTube over everything — blakesamic · 2026-10-12