Skill UI

Runs autonomous optimization loops on artifacts via Hypothesis Tree Refinement: persistent hypothesis trees, isolated git worktree experiments, evidence propagation, and a held-out merge gate to prevent overfitting. For model training, agent harnesses, data synthesis, and benchmarks.

Implements a comprehensive defense-in-depth framework for autonomous AI agents that interact with external tools. It mitigates risks like prompt injection or tool misuse by enforcing strict tool allowlisting, binding tools to least-privilege identities, requiring human-in-the-loop approval for high-impact actions, and utilizing policy enforcement (like NeMo Guardrails) to bound the blast radius and ensure compliance.

This comprehensive guide outlines best practices for building reliable and robust AI agent harnesses. It covers optimizing action spaces, defining structured tools, managing observation formats, and implementing essential error recovery contracts. Use these principles to dramatically improve agent completion rates, manage context budgets effectively, and transition AI prototypes into stable, production-grade autonomous systems.

Transforms Claude Code into a fully autonomous agent with persistent memory, scheduled operations, computer use, and task queueing. Replaces standalone agent frameworks using native crons, dispatch, MCP tools, and memory. Ideal for continuous automation and self-directed agent loops.

This pattern defines a robust, continuous loop structure for autonomous agents. It integrates crucial components like quality gates, evaluation harnesses, and explicit recovery controls. It guides complex workflows by defining decision paths (e.g., PR control, RFC decomposition) and includes strategies to mitigate common failure modes such as infinite loop churn or cost drift, ensuring measurable progress in autonomous systems.

Design and review goal-oriented agent loops to prevent failure modes like spinning or Goodhart-gaming. Covers defining machine-decidable goals, selecting loop types (servo/regulator), and implementing plan/build/judge skeletons. Essential for building stable autonomous agents.

Delegates coding, code review, and plan review tasks to OpenAI Codex CLI with sandbox controls and security boundaries. Supports autonomous execution, multiple modes, and safe review handling.

This tool takes vague, rambling descriptions of desired outcomes and transforms them into a single, highly polished, autonomous prompt. Instead of building the deliverable, it crafts the comprehensive prompt (including quality bars, tool mandates, and verification loops) that a fresh AI session can execute completely on its own. Ideal for structuring complex, multi-step AI tasks.

A comprehensive guide for authors targeting the International Conference on Autonomous Agents and Multiagent Systems (AAMAS). This skill helps authors strategically position their work by focusing on core themes like autonomous agents, multiagent learning, game-theoretic interaction, and decentralized decision-making. It covers submission framing, evidence requirements, and risk assessment to ensure manuscripts align with the venue's technical focus.

Generate decision-ready reports from AI agent evaluation runs. This skill ensures autonomous, assisted, failed, and timed-out outcomes remain distinct and comparable. It defines metrics with correct denominators, handles latency and cost populations honestly, and maps evidence to decision gates. Ideal for benchmarking, regression testing, and production readiness validation of AI agents.

Schedule AI agent runs safely with cron, systemd timers, external clocks, or built-in schedulers such as Hermes, covering heartbeat patterns, permissions, overlap locks, stop conditions and verification to avoid unsafe tight autonomous timers.

Turns your Kanban board (Asana, GitHub Projects, Linear) into an autonomous AI development pipeline with multi-worker Claude Code dispatch, deterministic quality gates, adversarial review, cost tracking, and recoverable task state.