agent-behavior

Standards for defining and evaluating agent behavior

The meta layer behind AI skills. Persistent behavioral overlays that make expected agent conduct explicit across changing tasks and capabilities.

Studying the gap between what agents know and when they act on it.

Give your coding agent a short break after real work, then measure what happens next. Universal agent skill & experiment harness.

Actonomy: A taxonomy for describing and analyzing agent behavior. Gao et al., NeurIPS 2026 — How to Interpret Agent Behavior. Official codebook, datasets, and trace analysis tools (ACT*ONOMY / Act-onomy).

Behavior-management plugin for DeepSeek Harness: tool-call discipline prompt section, failure-triggered parallelism convergence (pool drops to 1, auto-restores), consecutive-failure user intervention. Complements dsh-token-optimizer; works standalone.

⚡ High-frequency (~60Hz) in-browser behavior tree execution engine for autonomous AI agents. Reactive priority interrupts, safety guardrails & deterministic action replay over MCP.

The senior dev, unbundled — ten opinionated skills that make an AI coding agent push back, verify, attack its own code, and finish like a senior engineer.

High-performance routing engine that selects the best agent skill for a task and emits structured handoff decisions.

text systems, behavioral contracts, AI protocols, and small tools for keeping AI-assisted work operational. I built these tools to reduce drift, force useful decisions, and keep messy AI-assisted work usable.

pua — 挫败驱动的行为修正教练 skill：绩效文化话术驱动 agent 穷尽方案、拿证据闭环；失败升级、方法论路由、门控循环。平静首请求不触发，遥测默认关闭

Composable behavior mutators for agent prompts - priority, routing, batching, topology

RL-style eval measuring intent/action divergence in frontier agents: model acknowledges a correction, then acts on the stale value anyway. 3 scenarios, 655 trials on claude-haiku-4-5, Sonnet 4.6, GPT-5.4, and Gemini 3.1 Pro Preview.

Toy 7. An elimination-filter landscape applying two structural constraints simultaneously to map which objective classes can persist under sustained optimization pressure — and which cannot. Includes a four-stage scenario engine and open-question frontier. Companion simulation for The Shape of What Does Not End — Series 2, Part 4.

Agent authority guardrails for Claude Code & Grok — stop yield-back; drive reversible work yourself.

Canonical home for AI Behavior Science research and the Founding Territory Paper

Produces auditable token-usage and cost reports from runtime evidence, normalized usage bundles, and repository-level report sets.

Toy 4. An AI-powered goal-chain classifier and invariance stress tester. Enter any goal, trace the instrumental chain to its terminal type, then probe whether the classification survives symbolic, blind, independent, and adversarial re-examination. Companion simulation for The Invariant Drive — Series 2, Part 1 of the Architecture of Thriving.

Execution layer for skill-dispatcher — runs multi-phase agent chains end-to-end with per-step telemetry and chain_id correlation

Lint AgentSkill folders before install, sync or release: YAML frontmatter, name vs folder, description length, dev-tree paths, model-run telemetry, README standard. CLI, /skill-lint command and GitHub Action.