Open Source Agent Tools: November 14, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Gnomon 1.0 Brings Git-Style Branch, Diff, Merge and Undo to PostgreSQL Data So Each Agent Run Gets Its Own Branch
Gnomon (bitsinfra/gnomon, PostgreSQL License) appeared on GitHub and Hacker News on November 13, with v1.0.1 published the same day. It is a PostgreSQL extension for versions 15 to 18, self-hosted or managed (the README lists RDS, Aurora and Supabase), that tracks table history and lets a session work on a copy-on-write branch selected by a function call or a connection string. A branch produces a row and field level diff, merges are three-way and type-aware (counts add, JSON merges by key, text by line) with remaining conflicts returned as rows, and a merge or a single transaction can be undone. For agents, gnomon agent create issues a database role, a branch and a connection string; merge rules such as a row limit, a table allowlist and passing SQL checks are enforced by Postgres, and queued side effects such as emails and webhooks stay on the branch until merge. The project also ships an MCP server, a CLI, and TypeScript and Python SDKs that give each agent task its own run. The repository was created on November 13 and has few stars so far, and commenters on Hacker News questioned how much real-world use it has seen and criticized the README as machine-written, so the feature claims are the project’s own. Repository Source
2. Langfuse Releases a Beta AI Gateway in Rust That Traces Model Calls From Coding Agents and Other Uninstrumented Code
Langfuse announced the beta of Langfuse AI Gateway on Hacker News on November 13 as its second product after tracing and evals. The gateway is a stateless Rust service that exposes native OpenAI (/openai/v1) and Anthropic (/anthropic/v1) endpoints: a client changes its base URL and API key, the gateway forwards the request with a provider key held centrally, and each call is recorded as a Langfuse generation with tokens, cost, latency and time to first token. The founders say the target is model traffic that SDK instrumentation does not reach, naming coding agents such as Pi, Claude Code and Codex along with internal tools and scripts. It does not translate between provider APIs and relays responses unchanged, a choice the team attributes to subtle agent failures caused by imperfect translation. A traceparent header places a call inside an existing trace, and prompt and output capture is opt-in. The post gives the code as MIT-licensed, a 22MB Docker image, under 3ms median overhead with tracing on and more than 1,000 concurrent streams on 2 vCPUs; these figures are the vendor’s own. It can be self-hosted with Docker Compose or Helm or used through Langfuse Cloud, and budget caps, model aliases and fallbacks, guardrails and PII redaction are listed as planned. Repository Source
3. Kiln Publishes Harness Tuner, an MIT-Licensed Loop That Uses a Coding Agent to Tune an Agent’s Tools, Prompts, Skills and Model Against Evals
Kiln posted Harness Tuner (Kiln-AI/harness-tuner, MIT) to Hacker News on November 13; version 0.5.0 has been on PyPI since November 11. Instead of optimizing only the prompt, it runs experiments on the whole harness of an agent defined as a Kiln task: it rewrites tool descriptions and responses, moves long instructions into skills that load on demand, adjusts model settings such as thinking level and temperature, and can attach or drop RAG, fine-tunes and sub-agents. Experiments form a graph that branches from earlier versions, each one runs only the evals it targets, and a candidate is checked against every eval before promotion. The data split is enforced in code: the tuning agents see only the training set, the validation set decides promotions and the test set is never read. The tuner’s own agents run on Claude Code (the default), Codex or OpenCode, and it requires the Kiln app on the same machine plus model provider keys; the bundled demo costs about $2 in API calls according to the README. The headline result in the README, 2.5 times fewer failures at one sixth of the cost on a production agent with more than 200 tools, is Kiln’s own measurement and has not been independently reproduced. Repository Source