Feed wire.
Latest dispatches on the wire — plus a short, dated view of the work in motion.
Now
Celln
Building an execution plane for agents: isolated cells that borrow verified tools rather than assembling mutable environments. The immediate work is making the real KVM path easy to install, inspect and prove.
Sympozium
Exploring a clean split between coordination and execution. Sympozium should own workflow, policy and shared state; Celln may become the smaller execution primitive beneath an AgentRun.
Questions I am following
- What does a useful persistent agent API require beyond one-shot jobs?
- Which capabilities should an agent request explicitly: tools, files, egress, time, memory?
- How do we make revocation and provenance operational rather than aspirational?
Updated August 2026 · read the Celln notes →
Latest dispatches
5 items · also available as an RSS feed
The Cell Cannot Break
We asked five DeepSeek agents to escape their sealed hardware-isolated cells. At every layer — build, kernel, filesystem, seccomp — the cell held.
The Agent Environment Should Be a Lease
Why I built Celln: an execution plane where agents borrow verified tools instead of assembling machines, and why it may become Sympozium's default substrate.
Post-Kubernetes Infrastructure for GenAI Workloads
Field notes on Modal's million-sandbox announcement, what it says about Kubernetes' assumptions, and the coming decoupling of coordination from execution.
The Sticky-Note Problem: Why Multi-Agent AI Is Broken at the Coordination Layer
Why every multi-agent system built today is passing sticky notes between people in different rooms, and what the MAST study tells us about fixing it.
We've been building AI agents wrong.
Why two million LLM agents produced zero collective intelligence, and what a synthetic membrane between agents would look like.