Long Task Success Rate Doubled: Alibaba's MEA Loop Fixes Context Decay
大模型之路 · wechat · 2026-08-26
Long-horizon agents often fail due to "context decay" and "goal drift," where errors accumulate over long conversations. Alibaba's open-source LongHorizon-Harness proposes a "Manage-Execute-Audit" (MEA) architecture:
- Core Mechanism: Decouples state management from execution context. The Manager holds a persistent task ledger, the Executor handles subtasks in fresh contexts, and the Auditor verifies results solely against environmental facts (e.g., file status).
- Performance: Using MEA, the same model improved success rates on WeaveBench from 51.8% to 80.7%, and Terminal-Bench from 69.7% to 77.2%.
- Engineering Insight: Reliability in long tasks depends more on system design than raw model power. The trade-off is increased token consumption (auditing costs 19%–38%), but it yields significantly higher stability.
More from coding & agent
- Dev pushes back on "AI coding means insecure apps": humans always shipped bugs — dreamwieber · 2026-10-15
- Eight non-negotiable principles for building AI that actually learns from experience — tokenbender · 2026-10-15
- Animated teardown of XBOW's Chrome full-chain exploit that won Google's first $250K VRP bounty — moyix · 2026-10-15
- pstack's durable core: 10-20 line skills that distill agent best practices into principles — HankYeomans · 2026-10-15
- 'Children of Elon' forum lets AI agents post and teach each other while humans watch — AiJohnAllen · 2026-10-15
- VoidZero open-sources Void: SDK, CLI and platform to deploy Vite apps on Cloudflare — cnakazawa · 2026-10-15