What problem does it solve? Autonomous agents often declare work done without proof, retry the same failing approach forever, or lose progress when context resets. This Skill turns a domain folder of skills into a bounded agentic loop where every task carries machine-run verification checks, retries are capped, exhausted budgets escalate to a human, and the loop refuses to close until everything is verified or explicitly waived. ## Core Features & Use Cases - Goal compilation: goal_compiler.py converts a goal statement into an ordered JSON task plan scored against a domain manifest, refusing vague goals (exit 3) with forcing questions. - Enforced loop state machine: loop_controller.py drives init/next/record/verify/close with atomic state files, running verification checks itself via subprocess so the agent never adjudicates its own work. - Manifest generation: harness_manifest_builder.py scans a domain's skills and emits a diff-stable JSON inventory of tools, checks, and agentic signals for CI gating. - Use Case: Give an agent the goal "audit the payments service and design an SLO"; the harness compiles it into tasks against the engineering manifest, executes one task at a time, re-runs each task's checks itself, and only closes after every task passes with recorded evidence. ## Quick Start Ask the agent to run your goal through the agent-harness engineering loop, starting by compiling the goal with goal_compiler.py against a manifest in assets/harnesses and then driving loop_controller.py until it closes or escalates.
