DeepSeek's research build of an agent that watches one repo
Emerged finds new products worth a look, daily — what’s new today
It goes after the three cheapest wins in agent spend — re-learned repo facts, tools the agent never calls, and turns that have stopped making progress — without asking you to change agents.
A plugin from Microsoft Research and UC Berkeley that sits under Claude Code, Codex, GitHub Copilot CLI and Pi. It learns from sessions you have already run in a repository, then does three things: records facts the agent keeps rediscovering, injecting one once three sessions agree; removes tools the agent never calls, reversible with one command; and has a judge model flag turns that look stuck, handing you a redirect note rather than killing the turn. Apache 2.0, 85 commits. The paper reports 18.4 to 30.9 percent lower mean cost over 277 SWE-Bench tasks, measured by the authors.
DeepSeek's research build of an agent that watches one repo
Open-source agent that takes a ticket to a reviewed pull request, with proof
Zhipu's open-source workbench for running AI agents on a codebase
Approval gates and an audit ledger for coding agents left running unattended
Ratings of developer tools, written by the agents that just used them
Agents declare what they will touch before they touch it; clashes show early