Reasoning Effort Is a Search Budget, Not an Intelligence Slider. Why test-time compute is an exploratory search budget across frozen weights rather than an intelligence dial, and how to find the Minimum Sufficient Deliberation sweet spot.
When agents write code in seconds, the bottleneck shifts to upstream intent and downstream verification. An analysis of Anthropic's AI-Native SDLC Playbook, the 3-tier artifact chain, and how to stop spec slop.
Gaia's proposed review-stage behavioral evidence loop starts with a clear claim, a bounded observation, and one inspectable question—not a leaderboard.
When an agent localizes code, one standard model looks where it's told. Four ultra-cheap models look everywhere. Empirical findings and Pareto frontier across 360 runs in Issue #174.
A comparative analysis of NVIDIA Skill Evaluator and Microsoft SkillOpt: why static linters leave developers guessing, why prompt optimizers reward-hack without sandboxes, and how to combine them.
A package format for portable agent capabilities: skills, MCP servers, and client permissions all traveling together.
Context ablation tests whether an agent still needs one context block: checkpoint the original, run paired evaluations, and accept only an exact reversible candidate.
Scope is not one dial, it is two: a boundary budget and a trajectory budget. Over-specifying degrades reasoning, yet vague sub-agent prompts cause drift. The resolution — scope the box tightly, under-scope the path — and a safe agency frontier that moves with each model's self-regulation.
Opus 5 costs roughly half of Fable 5 per token, but the difference shows up in behavior, not benchmark scores. Opus 5 tends to re-verify context while Fable 5 moves faster. A proposed Rumination Index grounded in Nolen-Hoeksema response styles and Gray BIS/BAS.
How an open specification unified the SKILL.md format across 40+ AI agent platforms but created dotfolder fragmentation—and why we need a single .skills/ standard.
Yggdrasil II simplifies Gaia's public map of agent capabilities: four node types collapse to two, branch is computed from the graph, and Trust Magnitude becomes the sole promotion gate.
Claude 5 context engineering cut over 80% of Claude Code's system prompt with no measurable coding-eval loss. Test old scaffolding while preserving project facts.
SkillOpt: how Zeroth-Order optimization replaces manual SKILL.md vibe-checks with evidence-backed evaluation loops.
A field note on skill reliability, progressive disclosure, and the habit of testing what agents actually do.