Automated experiments that rewrite an agent's tools, not just its prompt
Emerged finds new products worth a look, daily — what’s new today
If your agent fails the same way every week, this reads the failures across runs and proposes changes to its harness that have to clear an evaluation before they stick.
GEAR runs a task agent against a benchmark, then hands the failures to a meta agent that looks for patterns across runs rather than inside one. What comes back are concrete edits to the harness: prompt wording, tool definitions, skills, how context is managed. Changes are promoted only when evaluation says they helped. It installs from npm and is driven from Codex or Claude Code through a bundled Refine Skill. MIT, 383 stars, 275 commits, no tagged release. The headline figure, a marketing task going from 27 to 40 percent, is self-measured, and the preprint lists arXiv as coming soon.
More like this
Skill packs that aim a coding agent at a compiled binary
Open-source infrastructure for agents that keep improving from use
Approval gates and an audit ledger for coding agents left running unattended
A home your team runs itself for its skills, MCP servers, hooks and prompts
Open-source agent that takes a ticket to a reviewed pull request, with proof