hi friends, Flux maintainer here. Stefan gave this talk at FluxCon EU
He maps out some "levels of detachment" for software & platform engineering that we've all been watching happen. Copilot autocomplete -> chat interfaces -> spec-driven development -≥ agent teams -≥ fully autonomous dark factories. Same progression applies to ops: agents that go into your clusters, diagnose issues, and present resolution plans without human intervention.
GitOps still matters in this world of agents:
1. Git as the interface to production (agents open PRs instead of running kubectl directly)
2. Git as persistent memory (incident RCA gets committed so the next similar event resolves faster)
3. Desired state that includes agent definitions (recover your AI orchestration the same way you recover a cluster: point Flux at a repo)
He also announced fluxcd/agent-skills, which are OCI packaged, verifiable skills for Flux that should progressively discovery instead of loading everything into context. We've seen a 32% improvement on knowledge tasks and 29% on repo audits vs. baseline models running against real public repos.
Stefan's anecdotes on developing agent tooling are relevant beyond Flux.
Every harness defines agents differently (markdown, TOML, JSON), there's no portability, and reproducibility and measurement is expensive and time-consuming.
stealthybox•2h ago
He maps out some "levels of detachment" for software & platform engineering that we've all been watching happen. Copilot autocomplete -> chat interfaces -> spec-driven development -≥ agent teams -≥ fully autonomous dark factories. Same progression applies to ops: agents that go into your clusters, diagnose issues, and present resolution plans without human intervention.
GitOps still matters in this world of agents: 1. Git as the interface to production (agents open PRs instead of running kubectl directly) 2. Git as persistent memory (incident RCA gets committed so the next similar event resolves faster) 3. Desired state that includes agent definitions (recover your AI orchestration the same way you recover a cluster: point Flux at a repo)
He also announced fluxcd/agent-skills, which are OCI packaged, verifiable skills for Flux that should progressively discovery instead of loading everything into context. We've seen a 32% improvement on knowledge tasks and 29% on repo audits vs. baseline models running against real public repos.
Stefan's anecdotes on developing agent tooling are relevant beyond Flux. Every harness defines agents differently (markdown, TOML, JSON), there's no portability, and reproducibility and measurement is expensive and time-consuming.