Clone-and-run loops from the research stack

SWE-agent

The 100-line AI agent from the Princeton SWE-agent line. Research origin, radical simplicity, still beating bloated wrappers on SWE-bench-class work.

Use caseUse it to throw a GitHub issue at a 100-line loop you can actually read, fork, and instrument.

HKUDS

Open agentic coding from HKU: agent harness, loop engineering, multi-agent orchestration. Academic lab software that ships as a repo you run.

Use caseUse it when you want a research harness with a real coding loop, not a chatbot with extra steps.

e2b-dev

Computer-use agents on open models plus an E2B desktop sandbox. The open alternative to a vendor's computer-use API.

Use caseUse it to let an open model click through a desktop app inside a throwaway VM.

lastmile-ai

Build effective agents with Model Context Protocol and small workflow patterns.

Use caseUse it to compose MCP servers into a durable agent without standing up a full IDE plugin.

awslabs

AWS Labs multi-agent orchestration for AI coding CLIs. The interesting part is the coordinator, not the branded CLIs it wraps.

Use caseUse it to run two vendor CLIs as workers under one policy instead of alt-tabbing between terminals.

exe.dev

A 700-point HN essay arguing that the tools you code with have to be inspectable. Agent CLIs are becoming the new IDE, and lock-in is the default path.

Read

xlang-ai

The long-horizon computer-use benchmark, version 2. If you are training or grading desktop agents, this is the eval the serious repos cite.

Shipped

provos

A secure runtime for autonomous agents with policy from plain-English constitutions. Aimed at people who will not give a coding agent root.

Read