Agentfork: fork an LLM agent mid-run and sweep the broken step
Show HN (Oct 17): agentfork records every LLM and tool call, then lets you fork a failed run at any step and sweep prompts, models, or tool results with pass rates instead of anecdotes. Earlier steps replay from a local cache so you only pay for the step under test; OpenAI/Anthropic wrappers plus LangGraph support. Headline demo: a refund agent that reads cents as dollars keeps failing on bigger models until one system-prompt line fixes it (3/3 pass). pip install agentfork-cli; traces stay on-disk in .agentfork/traces.db.
Save ideas in this browser. Upvotes count across the board.
Show HN / GitHub rebase-energy/hillclimb (MIT): open CLI + Python framework that uses Claude Code and Codex as search operators to autonomously propose Python programs maximizing a score you define—AlphaEvolve-style hillclimbing for speedups, optimisation, and prediction tasks. Modular strategies and evals aim to discover better autoresearch methods, not just wrap a single agent loop. Docs at docs.hillclimb.sh; PyPI package hillclimb (Python 3.12+).
EtherPK’s /Agent (Alt+A) opens a read-only agent chat inline in a markdown/outliner PKMS: you pick context scope (block tree, document, links, or whole-graph tools). BYO keys (Anthropic/OpenAI/Gemini/OpenRouter/Ollama); provider calls stay in-browser; only selected turns save back as outline blocks. Solves cold-start chat and lost answers by anchoring agents in durable notes.
MIT Python package signs every policy-gated tool call into a hash-chained .tkb file verifiable offline (Python+TS); works across Claude Code, OpenAI Agents, LangChain/LangGraph, MCP and more—agents can edit their own logs but not hide edits when the signer is out of reach.
Show HN / GitLab pablosanzo/musejailbreak: independent macOS menu-bar bridge (not affiliated with Meta) that tunnels a dedicated local Chrome to Muse so the personal agent can see the screen and click/type on sites where you are already logged in. Cloudflare Quick Tunnel with AES-256-GCM E2E (fallback sish); share URL+token with Muse after toggle-on. Explicit risk warning: real browser control including purchases—use only with trusted agents and never unattended on sensitive accounts.