Greenfield DevOps platform
- status
- paused · mixed
- when
- 2026-07
- came from
- Team enablement, First public repos, Agent Zone, Dream Team
- led to
- SnapDragon-Coder, Fleet & operations
An agentic DevOps platform on a home Kubernetes cluster: GitOps, workflows, a model gateway, review bots, and a telemetry-to-escalation-to-harvest loop.
Why it started
Because Dream Team could not ship. A soak on real backlog items merged 0 of 10. The 2026-07-02 plan replaced the org-chart-as-engine with one Temporal engine at every scale, git as the only write path to infrastructure, and Kubernetes as the substrate contract. The platform I would want at work: agents that review code, run canaries, and escalate to a bigger model only when a small local one is out of its depth, with the harvested escalations becoming the hard examples a small model is trained on next. Full arc: Rabbit Holing.1
The cutover was decided on evidence: six pipeline fixes took yield from about 0 of 9 to 40% with zero false merges on the same corpus; I signed the go on 2026-07-06 and the old cluster was deleted that day.2
What it is
k3s on carlo, ArgoCD for GitOps with encrypted secrets, Temporal for anything long-running, a LiteLLM gateway that puts local and cloud models behind one API with profiles named for jobs rather than models (the reasoning is on Local or cloud), and a set of agents and workers. The telemetry → escalation → harvest spine was built to work item 5 of 7.3
Architecture, as built
One level up on three of these boxes, with their own diagrams: how a model call is routed, how Temporal runs one item, and what an agent is inside its harness, on How the pieces work.
Everything custom is application code on five standard contracts: the Kubernetes API, Postgres, Temporal, a git-forge API, and OpenAI-compatible inference behind a gateway. The agents own nothing but the cognitive step inside an activity; the harness owns checkout, branches, pushes and state, so a worker’s identity never appears in an output. Alerts re-enter the same funnel as work items, which is the telemetry → escalation → harvest loop in one picture.4
Where it stands
Paused since 2026-07-25 by choice: the fleet went into “training mode” for Marilyn and every agent workload was scaled to zero. The restore recipe is written. The platform pieces that survived are the ones every later project stands on: the cluster, GitOps, Temporal, and the gateway.
Lineage
- Came from: Dream Team (the system it replaced), Agent Zone, first public repos, team enablement
- Led to: SnapDragon-Coder (the platform wanted a local coder), fleet ops
Sub-pages
- How the pieces work: the gateway, one item’s life in Temporal, and an agent inside its harness
Sources
Numbers on this page are quoted from the lab's own records. The records are private; each note gives the record's date.