#agent-harness

Finish-First Autonomous Agent Loop for Claude Opus 4.7 – 2026 Edition

An agent harness that compiles a model into one provably-correct, self-retargeting CUDA megakernel and self-tunes it past cuBLAS at batch-1 LLM decode, paper: https://arxiv.org/abs/2606.09682

Fable-style spec + evidence gate for Claude Code + Codex. Makes Opus/Codex work under Fable-like discipline: blocks every edit until a deterministic spec passes, and there is no "done" without live acceptance evidence. Spec-first, verification-gated, forbidden-paths enforced.

Skill repository to enforce Engineering disclipline in coding assistants

AI-assisted XAUUSD daily research workflow with Research Skill, Agent Harness, LangGraph, Tool Calling, and free public data sources.

An interactive roadmap for harness engineering — building the agent loop, tool layers, context engineering, memory, retrieval, eval harnesses, and the rest of the scaffolding around modern LLMs.

OS-level safeguards layer for AI agent loops — runaway-kill, timeouts & resource limits so coding agents (Claude Code, Cursor, Codex, Antigravity) in YOLO mode can't burn tokens or freeze your Mac

HarnessRouter Community Edition: the self-hosted, Apache-2.0 edition of the unified interface for agent harnesses. Run Codex, Claude Code, Hermes, and more through one API, with sessions, streaming, files, cancellation, and failure handling. Implements the Unified Harness Protocol (UHP), an open standard. Your keys, your infrastructure.

Local execution harness for OpenAI Codex with durable ledgers, verification gates, memory, and replay-gated skills.

A local-first coding agent: runs the heavy executor on your local model and escalates to the cloud only when stuck. Out-resolves single-shot Opus/Sonnet by planning, running the project's real tests, and retrying — at ~half the cost of an all-cloud chain. Graded by SWE-bench, not an LLM.

A lightweight local Coding Agent harness inspired by learn-claude-code s20, featuring a minimal LLM tool-use loop, workspace-safe file access, command execution, and pytest-verified Phase 1 tools.

A generator-critic harness for Claude Code: Planner / Executor / a parallel panel of tool-backed jurors / a routing Judge, with candidate isolation and CI as the final gate.

DeepSeek-first local Windows agent for files, research, automation, verified computer use, and auditable office workflows.

A lightweight, explicit, Git-native coding agent continuity and governance harness.

Terminal agent with durable sessions and background jobs, a fail-closed filesystem sandbox, and first-class unattended runs. Written in Go.