meta-context-engineering

MCE (Meta Context Engineering) — Ye et al. 2026, surveyed in 2026-07-07-harness-engineering-self-improvement. Automates context-engineering one level above ACE: instead of optimizing context content, it separates the mechanism (how context is managed) from the artifact (what is in context) and optimizes both via bi-level search.

The bi-level structure

• Inner loop — a base-level context engineer optimizes the context function $c_s$ for a given skill $s$.
• Outer loop — a meta-level agent evolves the skill itself, via agentic crossover over a skill database ${\mathscr{H}}_{k-1}$.

A skill defines a context function $c_s=({\rho }_s,F_s)$: static components ${\rho }_s$ (prompts, knowledge bases, code libraries) plus dynamic operators $F_s$ (search, selection, filtering, formatting). Implementation uses standard coding-agent tools — {Read, Write, Edit, Bash, Glob, Grep, TodoWrite} — i.e. the optimizer is itself a coding agent.

Contrast with ACE

| | ACE | MCE | |---|---|---| | Optimized object | Context content (playbook bullets) | Context mechanism + content | | Structure | Heuristic rules enforced (bullet format) | Free-form skills, no enforced structure | | Search | Reflect-and-curate loop | Bi-level: inner context, outer skill evolution |

Position in the wiki

• Sits between “structured context” and “harness code” on the recursive-self-improvement ladder; meta-harness goes one step further by optimizing the full harness program.
• One of Weng’s named open questions: whether MCE/Meta-Harness-style optimization stays stable across many iterations or hits diminishing returns / reward-hacking (2026-07-07-harness-engineering-self-improvement).

(2026-07-07-harness-engineering-self-improvement)