Beacon memory now supports @subquadratic alongside Jev, both via the System One decisions API! Pick your evaluator with one flag, and Beacon turns the sessions that worked into lessons every agent reuses over MCP! Github link in comments 👇

~5 seconds to create your own profile! Let's see if anyone has higher Codex usage than @fkodom... (+ you get free @asymptotelabs merch!)

AI agents are generating an unprecedented amount of runtime data, yet we're throwing away most of its value! Every agent session reveals what worked, what failed, and where time was wasted. But as agents take on more complex work, extracting those insights from raw logs becomes nearly impossible. Introducing Beacon Lenses 🔍 Turn any agent trace into an interactive tool to: → Debug failures → Inspect code changes → Identify bottlenecks → Analyze costs → Investigate security risks Beacon ships with 7+ built-in lenses, but you can build your own for any use case. Just tell your coding agent what you want to explore, and it can generate a custom lens you can use directly in Beacon Cloud. Your agent traces aren't just logs anymore. They're building blocks for entirely new tools. Start building: beacon.sh Docs in the comments 👇

We've been keeping our cards close to our chest. In less than two weeks, @asymptotelabs is going all in on our most ambitious bet yet. 🃏 Stay tuned. The reveal is coming.

Another insane Beacon use case! Your coding agent hits a context limit, a rate limit, or a wall. You switch tools… and start explaining everything from scratch. Beacon by @asymptotelabs fixes that with handoffs: move an in-progress task to any other harness, with the context that actually matters. Handoffs work across all the major harnesses: →Claude Code →Codex →Cursor → OpenCode, and 20+ more GitHub repo: github.com/Asymptote-Labs/ag… (don't forget to star it ⭐) Beacon already captures your full session history across harnesses. But dumping a raw transcript into a new agent isn't a handoff. It's noise. Most sessions are full of dead ends, failed commands, and stale assumptions. The next agent doesn't need all of it. It needs: →what the task is → what's been tried and ruled out → what changed in the code → what the human corrected along the way → what's left to do Beacon distills the session into exactly that, and any harness can pick up where the last one stopped. In the demo video, Claude Code hits its limit mid-fix. I hand off to Codex, which already knows the goal, the failed approach, and my correction, and finishes the job. Hand off to any harness. Your work shouldn't be locked inside one agent.