#Harness
Anthropic measured what actually works in AI agent memory: four layers, telling the agent what memory is for instead of rules, and dreaming, an offline step where the agent rereads transcripts and fixes its own memory. CL-Bench, Pokémon and wiki vs search numbers, plus Dreams API code and how to apply it in Claude Code.
55% of the improvements an agent proposed for its own harness fixed one failure and broke something that worked. A PHP tutorial for a self-improving harness with the piece that makes it engineering: the regression gate. Git-versioned rules, failure mining, minimal rules, protected cases, provisional promotion and pruning.
Across 54 cycles, the agent said it improved every time. In 56%, nothing changed or it got worse. The progress mirage measured in a paper, why a bigger LLM judge doesn't fix it, and how to build a Laravel loop with an external verification gate: checkable exit, hard budget, and a guard that stops the agent from editing its own exam.
An agent that runs across several context windows forgets everything each session. The pattern Anthropic, Manus and LangChain converged on: move state out of the window and into files. A JSON feature list that only accepts status changes, a progress log, a session start ritual and offloading large results with a restorable pointer, all with PHP code.
A subagent isn't a job title, it's a context window. When to use subagents (and when one agent is enough), the difference between isolated and fork modes, the numbers from Anthropic and Cognition, and how to implement both in a PHP harness straight on the Claude API, no framework.
The harness that saved your agent last year is the one getting in its way now. Every model generation turns the previous harness's crutches into noise: ALL-CAPS rules, redundant tools, giant few-shot blocks, rigid orchestration. What the official Anthropic and OpenAI guides tell you to remove, Rich Sutton's lesson applied to harnesses, and how to prove with evals that removing things didn't make anything worse.
Your CLAUDE.md is 400 lines long and Claude ignores half of it. A tutorial on how to use skills in Claude Code: folder structure, SKILL.md frontmatter, progressive loading, scripts and references inside the skill, how to test whether the agent triggers it, and the decision matrix between skill, MCP, CLAUDE.md, slash command and subagent. Built on the real skills that run this blog's pipeline.
On Oct 22, 2026, Anthropic donated the Agent Skills standard to the Agentic AI Foundation, the same path MCP took. What agent skills are, how SKILL.md progressive loading works, an audit of the 8 skills that run this blog, and a checklist for the skills you already have.
Google open-sourced Harness2, a framework where the agent rewrites its own harness (prompt, tools, memory, skills, subagents) without touching the weights and without ground truth. What it edits, what the numbers actually prove, how to run it, and how to govern a harness that edits itself.
Higgsfield MCP hands your agent more than 30 image and video models, with the generation harness already built. How to connect it to Claude Code, what each generation costs in credits, and 10 systems where a dev can integrate it.
An artisan command, a queue and Claude Code in headless mode (claude -p) generating 11 sites inside the subscription. How to orchestrate the agent from PHP, where the math works out, and how to plug in Higgsfield as an asset step with a budget controlled by code.
Grok 4.7 landed on September 21, 2026 costing 5x less per token than Fable 5.1 and GPT-6 Astra, and it still loses to Astra on cost per completed task. Where this model actually pays off (spoiler: latency, not code), where it sits 22 points behind on Terminal-Bench, and the math that takes apart the list-price comparison.