OpenHands 1.0: a self-hostable coding agent hits 68% on SWE-bench
Redesigned as a composable, sandboxed SDK you can run on a frontier API or fully local on Devstral — plus a Claude Code security fix and Copilot's agent team.
Copy markdown
Self-hostable, and not a toy
OpenHands 1.0 scores ~68% on SWE-bench Verified (500 real GitHub issues) driving a frontier API, and ~46.8% running fully local on open-weight Devstral 24B — matching Devin 2.0's public number. Install via Python 3.12+/uv or a standalone binary.
Security is first-class now
The rewrite is event-sourced: every prompt, command, file change, and error is an immutable log entry, so runs replay deterministically and recover cleanly. A SecurityAnalyzer risk-rates each tool call, and a ConfirmationPolicy gates high-risk actions for your approval.
Build this weekend: an overnight PR bot on your own box
Point it at a repo with SANDBOX_NETWORK_DISABLED=true and the confirm-risky policy on, and it can triage your issues and open PRs while you sleep — no API bill and no code leaving your machine on the Devstral path.
For coders: patch Claude Code now (2.1.267)
The Sept 9 release fixes a plugin-marketplace path-traversal vulnerability on macOS/Linux, adds a maxEffortLevel setting to cap effort and spend across Bedrock, Vertex, and Foundry, and adds --system-prompt-snapshot off to iterate on prompts without busting your cache.
Elsewhere: Copilot runs an agent team
GitHub's new multi-agent workspace puts specialized agents on implementation, testing, and docs at the same time, sharing one context window across the codebase instead of one assistant doing it all.
Heads-up: Google says attacks went agentic
Google's threat-intel group reports attackers shifting from single prompts to automated agentic chains that plan, execute, and iterate — compressing detection windows. If your own agents hold tool access, that new ConfirmationPolicy stops being optional.