Open-source agent that takes a ticket to a reviewed pull request, with proof
Emerged finds new products worth a look, daily — what’s new today
Harness Tuner
Automated experiments that rewrite an agent's tools, not just its prompt
Is this yours? Claim itMany agent failures sit in the tool layer, and this automates the tedious change-and-measure cycle while keeping a held-out test set it never sees.
Harness Tuner, from Kiln AI, improves an agent by changing everything around the model: it rewrites tools, trims contradictory prompts, moves long instructions into skills, and tries other models and settings. You state goals such as quality, cost or speed in plain English. It works in waves of single-change experiments, scored on a training split and promoted only on validation. The test split is kept from it, so the final number is yours to run. You need Kiln, evals, and Claude Code, Codex or OpenCode. MIT, installed with uv.
More like this
An open-source safety layer that checks every AI agent action before it runs
An idle game where the model never improves and you build the scaffolding
A local ledger of every tool call your coding agents make
Open-source infrastructure for agents that keep improving from use
A local message bus that lets your coding agents review each other's work