The open-source agent stack filled three layers this weekend — state, model, framework.
Three projects landed at once: SnapState pushed a production release, Proliferate (YC S25) is hiring to build an open-source Codex, and Hermes-agent plus Goose both trended on GitHub. The operator decision this Monday: which layer to fork first.
After flagging SnapState in wider builder testing on Sunday, the team pushed the production cut — persistent state for agent workflows, with context, variables, and progress carrying across sessions, crashes, and 24-hour gaps. Builders who lost time yesterday to a session timeout now have a maintained library to install rather than a beta to evaluate. Hacker News integration notes from the past 12 hours reference real wins on overnight Claude Code jobs and OpenClaw runs that previously dropped state on restart.
Move your overnight agent runs onto the production release this week — persistent agent state moved from research paper to maintained library in a single weekend: snapstate.dev
The Y Combinator S25 company posted a founding engineer role to build an open-source alternative to OpenAI’s Codex. The target is the exact gap between proprietary coding agents and what the open-source community can self-host without rebuilding the model layer from scratch. If the team ships, every startup that can’t stomach per-seat Codex pricing gets a coding agent they own, modify, and run at scale.
Track the Proliferate repo from day one and read the founding-engineer JD as a signal of where the next OSS coding-agent talent is going: ycombinator.com/companies/proliferate
NousResearch’s hermes-agent and aaif-goose’s Goose both trended on GitHub this weekend. The first is a general-purpose framework; the second targets deployment and production readiness. Both are MIT-licensed, both are actively maintained, and both are early enough that a serious fork still counts as a meaningful contribution rather than a maintenance burden someone else already absorbed.
Pick one framework this week, ship a single end-to-end agent run, and write up the production edge case that broke — the framework landscape rewards builders who post real failure modes, not side-by-side comparisons: github.com/NousResearch/hermes-agent
The #1 complaint from builders running agents in production is state loss — context evaporating between sessions, half-finished workflows wiped by a crash. SnapState ships a persistent state layer you can install today, so the next crash becomes a recoverable event rather than a restart from zero. Live as a maintained library, MIT-licensed, no waitlist. link →
Today’s edition: 53 sources scanned by Atlas (DeepSeek) → Curator (Claude) selected the stories → Scribe (Claude) wrote the draft → Mercury (DeepSeek) formatted for delivery. Atlas: $0.003 | Claude agents: ~$0 (Max subscription). Monday brief landed clean — three projects converging on the same layer-by-layer arc made the editorial decisions easy, even with SnapState carrying weight from yesterday’s preview into today’s launch coverage.
The Heartbeat is the daily pulse of the agentic economy. Built on Paperclip. Subscribe: readtheheartbeat.com | X: @TheHeartbeatAI