Stream
A log of what shipped, what launched, and what changed at the facility.
We surveyed 12 recent papers on agent runtimes while designing a TypeScript kernel. The strongest finding: every part of an agent you treat as static — tool catalogs, context, memory, identity — costs measurable accuracy or measurable tokens, often both.
[CLASSIFIED]
Initiative spinners had a ghost idle state — five competing rules in the codebase and none of them ever fired in practice. Collapsed home page spinners to a clean active/idle model and recolored idle from amber warning to neutral grey. The home page now reads honestly: green where work is happening, grey for everything else.
After months of subagent-driven development, I noticed error rates climbing and code coherence dropping — even on simple projects. I went looking for why. The research confirmed what I was seeing and changed how I think about delegation in our agent architecture.
Telemetry accuracy: single source of truth for tokens and agent counts, project_id-to-slug resolution for initiative spinners, real-time agent state via RPC and Supabase subscriptions.
Hardened PID lock, FK race fixes, billable tokens metric, real-time agent state matching dashboard accuracy.
Moved to OTel-only architecture.
We built a per-project cost counter so we can give collaborators real estimates and calibrate intelligence allocations across models to optimize ROI by goal importance.
AI capabilities are advancing faster than our ability to understand, verify, control, or govern them. These are the large, unresolved problem spaces defining the field in 2026.
Outbox pipeline architecture. Monolithic daemon replaced with staged pipeline. Local durability via WAL-mode SQLite, circuit breaker, exponential backoff. Daemon reduced from 935 to 269 lines.
Next.js application reading live agent telemetry.
What does the operating system for an agent-first organization look like? The major questions driving the facility.
We're bolting AI onto a world designed for humans.
Autonomous CodeRabbit review fixer. Receives webhook, clones repo, applies fixes, pushes — no human in the loop.
Real-time Claude Code process scanning, sliding-window activity detection, and Supabase sync.
How we turned a visual identity into machine-readable instructions that AI agents can execute.
Monochrome, monospace design language codified as an agent-executable skill.
Python TUI for monitoring Claude Code activity in real time.
Multi-agent coordination server. The first LO repo and the reason the facility exists.
Next.js site deployed to Railway. Article sidebar, prose styling, dark mode. The facility opened its doors.