As models commodify, measured performance and cost are migrating to the harness/context layer—orchestration, execution-state design, and context lifecycle—making ‘context engineering’ the decisive engineering variable.
Evidence
- ‘Same Model, Different Harness’ and ‘The Harness Is the Thing’ show harness design and planner/worker/critic splits drive results independent of model capability.
- Structured execution-state approaches (SKILL.state, Scroll’s context-as-environment, agentic context lifecycle primitives, Weighted Memory Tree’s 32.8% token savings) replace ever-growing history to fight context degradation.
- Empirical usage studies show agents consult instruction/working-note files 6x more than formal docs, and first-chunk inclusion matters more than within-chunk ranking.
- Efficiency architectures (GLM-5.3 Flash at 1/40 cost, GQA+sliding-window energy findings, prompt-repetition plateau after 4x) show cost/quality tuned via architecture and context rather than model choice.
Implications
- Competitive leverage and spend optimization now lie in harness/context design—frontier models for planning, commodity models for execution—more than in picking a model.
- Documentation and tooling should be restructured for agent consumption (AGENTS.md-style, minimal opinionated prompts) to capture these gains.
Concepts
Context Engineering Code Agents Agents Open Source LLMs Tool Use
Confidence
medium