v4.8.0 — Context Engineering Improvements

What's New

Four improvements from Agent Skills for Context Engineering research.

Justification-Before-Score

Evaluator agent now reasons through correctness BEFORE assigning a score. Research shows this ordering improves LLM-as-judge reliability by 15-25%. Also added position bias warning.

--format summary (Observation Masking)

read_results.py and trace_insights.py now support --format summary — compact output (~200 tokens vs ~5K full JSON). Returns aggregated metrics, failure pattern, and top 3 failing inputs. Prevents context bloat when loading analysis data into proposer strategy.

Strategy.md Filtering

Strategy documents now contain ONLY current iteration data — no stale info from prior runs. Research shows a single irrelevant document causes disproportionate degradation (step function, not linear). Top 3 promoted memory insights only. 1500 token cap.

Anchored Iterative Summarization

Consolidator agent treats promoted insights (rec >= 3) as immutable — they're never re-summarized or paraphrased. New observations extracted from proposal.md use literal text to prevent telephone-game fidelity loss. Garbage collection removes observations not recurring in 5+ iterations.

Install

npx harness-evolver@latest

References

• Agent Skills for Context Engineering
• NousResearch/hermes-agent-self-evolution

🤖 Generated with Claude Code