If Everything Can Be Recreated, Who Gets Remembered?
When AI agents make software implementation reproducible on demand, original authorship does not disappear. Why Gaia Skill Tree and Yggdrasil preserve verifiable lineage and evidence-backed trust.
Blog
Blog posts
Capability Amplification: When Better Scaffolding Beats a Bigger Model
Structured agent scaffolding can buy task-level reliability more cheaply than buying a larger model. Why procedural skills are an economic instrument rather than prompt debt.
The Return of the Simple Reflex Agent: Why Jev Is a Classifier, and Why That Matters
Jev makes fast typed decisions instead of generating prose. Its useful return to reflex agents and probabilistic classification puts frontier LLM calls and local classifiers on the same audit list.
Why Your Multi-Agent Setup Costs More Than a Single Heavy Agent
Cache goes cold every time a subagent takes over 5 minutes. Your orchestrator re-reads its own 100k context at 12.5x warm rates, and you never see it on the invoice.
We Ran 48 Coding Sessions to Find the Real Compaction Sweet Spot
Phase 1 predicted the sweet spot was 40k–65k. We ran 48 live coding sessions across 8 autocompaction thresholds on Gemini 3.8 Flash. Here is the true Pareto frontier.
The Context Compaction Curve
You're at 100k tokens cold. Compact now or keep going? The cache-TTL economics, fresh-session alternatives, and the rule of thumb that should replace gut feel.
Don’t Make the Model Think Harder Than the Problem
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.
INTENT.md: Harness or Process Theater?
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.
Arbor Opens Its Evidence Loop for Review
Gaia's proposed review-stage behavioral evidence loop starts with a clear claim, a bounded observation, and one inspectable question—not a leaderboard.
The $0.003 Scout Fleet: Why Four Parallel Cheap Models Beat One Expensive One
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.
Evaluator vs. SkillOpt: The Gatekeeper and the Tuner in Agent Skill Engineering
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.
Agent Plugins: The Skill Landscape Just Got a Package Format
A package format for portable agent capabilities: skills, MCP servers, and client permissions all traveling together.
Context Ablation: Press Delete Without Losing the Experiment
Context ablation tests whether an agent still needs one context block: checkpoint the original, run paired evaluations, and accept only an exact reversible candidate.
Constrained Autonomy: Why a Flawless Brief Can Make Your Sub-Agent Dumber
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 vs. Fable 5: Rumination, Overthinking, and the Hidden Cost of Being Half the Price
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.
The Minimalist Irony of SKILL.md: Format Unity, Six Dotfolders, and the Case for .skills/
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: The Skill Tree Stops Storing What It Can Compute
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: Why a Smarter Model Wanted a Shorter Prompt
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: Zeroth-Order Optimization for Agent Skills
SkillOpt: how Zeroth-Order optimization replaces manual SKILL.md vibe-checks with evidence-backed evaluation loops.
Don't Ship Skills Without Evals
A field note on skill reliability, progressive disclosure, and the habit of testing what agents actually do.