I build the tooling people run when the cluster is on fire, and the layers agents need so they don't set it on fire again.
I started with a terminal, worked on MicroK8s at Canonical, ran the platform and SRE teams behind services you use every day, founded K8sGPT and took it into the CNCF, shipped Rust command-line tools like llmfit, talked at KubeCon on both sides of the Atlantic, wrote a position paper on the synthetic membrane, and keep on shipping.
Field record
010GitHub stars · all repos
02CNCFK8sGPT · founder
03KubeCon + CloudNativeCon North America 2023
040projects maintained
05Synthetic Membrane paper v2.2 · September 2026
06Rustoperators for Kubernetes · PlatformCon
07UKmade in
08Blueprint · systemsdesign.site live product
09Rust · Go · Python field tested
Workbearing000°fix-- / 10strength·····
open source · ten fixes hover a line to take a bearingtap a line to take a bearing
Findings
what I have been thinking about · what came out of it
One question, worked on in the open: how should AI agents share what they know, and who gets to act on it? The paper is the current answer. The figures are the evidence it rests on. The ledger is how it got here.
v2.2current revisionSeptember 2026position paper
The Synthetic Membrane: a coordination layer for multi-agent AI systems
A coordination contract for shared evidence, selective access and action ownership between agents, with explicit limits and an evaluation plan. The claim is bounded: if a controlled blackboard matches the membrane with lower overhead, the separate abstraction has not earned its place.
hypothesisBounded, with a prior-art comparison against protocols, frameworks and the blackboard.
contractAuthority and consistency rules for who may read, contest and act on shared state.
worked exampleA suspected ransomware incident, run through the contract with its limits stated.
evaluationMatched baselines, cost measured at the model boundary, and mechanism ablations.
evidence · seven figures · tap to openpaper figures in ink · field notes on cream
revision ledger
v2.2 Revised position paper.
bounded hypothesis and prior-art comparison
authority and consistency contract
worked incident scenario with explicit limits
matched baselines and mechanism ablations
Sep 2026 · current
v2 First revision under the coordination-layer framing: the membrane as a contract rather than an architecture.Jul 2026 · superseded
v1 Original position paper proposing a six-layer membrane architecture. Withdrawn from the site. The companion note, agents need somewhere to share their work, still stands.Apr 2026 · withdrawn
05Testing Celln's BoundaryFour boundary probes and one successful computation: what a small Celln demonstration shows, and what it leaves untested.Aug 2026 · note
04The Agent Environment Should Be a LeaseWhy I built Celln to lend agents a defined capability set, and what that execution model still needs to prove.Aug 2026 · note
03Post-Kubernetes Infrastructure for GenAI WorkloadsField notes on Modal's million-sandbox announcement, what it says about Kubernetes' assumptions, and the coming decoupling of coordination from execution.Jul 2026 · note