skyflo

Engineering log44 entries

The engineering log.

Build notes from the people making Skyflo: what shipped, what we cut, and what the work taught us.

Notes on AI coding agents, orchestration, multi-repo engineering, and persistent memory.

Latest entryOct 03 2026·4 min read

Start Where the Work Will Run

A mission can begin in Skyflo Cloud without cloning its repository on your Mac first. The useful boundary is the mission: its instruction, branch, work and conversation stay together wherever it runs.

Read the entry

The engineering harness

Agent orchestration, multi-repo missions, model routing, and engineering memory.

2026

14 entries
Oct 02The Mission Outlives the LaptopSkyflo 1.5 introduces Skyflo Cloud: a mission moves from your Mac with its history and uncommitted changes, keeps working while the Mac sleeps, and comes back when it is idle. How the handoff keeps one owner at a time, what the machine holds and what it can reach, and what the work costs.11 minOct 01Approval Belongs to the CommandAn agent can understand your permission and still reach the wrong execution path. Skyflo 1.5.1 carries the mission's approval policy to the command and file change itself, while messages from other missions keep their own identity and limits.4 minSep 29A Message Is Not the PersonSkyflo 1.5 lets missions on one Mac message each other, so one coding agent can ask another a question without a person carrying it between windows. The work was deciding what a message from an agent may and may not do.8 minSep 26What a Sandbox Has to ProtectSkyflo 1.4.2 fixes three things about the sandbox its agents work in. None of them was an agent writing outside its area. Each was about what that area touches: files the agent never saw, settings another program runs, and tools the person already relies on.6 minSep 26An Approval You Can ReadSkyflo 1.4.3 changes what a paired phone shows when an agent asks for permission. Testing on a real iPhone, we found it could ask for Face ID on a command it never showed. A signature proves who said yes. It cannot prove they knew what they were agreeing to.5 minSep 25The Phone Decides, the Mac ActsSkyflo 1.3 lets a phone follow your missions, answer their questions and approve their actions. The design question was never how to put an agent on a phone. It was how to let a phone say yes without letting it do anything.8 minSep 25Take Over, Hand BackSkyflo 1.4 lets you and an agent share one browser and one terminal. The hard part was not letting a person grab the mouse. It was deciding who owns the surface at every instant, what the agent may see of your turn, and where the mission resumes when you give it back.9 minSep 23A Voice Is a Control SurfaceSkyflo 1.2 lets you talk to a mission while it runs, and the voice drives the product rather than describing it. The hard part is not speech recognition. It is deciding, sentence by sentence, whether what you said was work or a control.10 minSep 18Skyflo Improves How It ImprovesSkyflo 1.1.0 is the release where the engineering harness becomes a HyperAgent. What that means, the three layers it is built from, and the boundary that keeps a system which changes itself accountable.11 minJul 16Why AI coding agents need an engineering harnessA category thesis for the layer around coding agents, with current Skyflo behavior separated from the complete architecture that remains a target.12 minJul 08One Model Is the Wrong NumberEngineering roles have different model needs. The architecture case for per-role routing, the provider access Skyflo supports today, and the managed lane that remains Planned.9 minJun 28The Case for Engineering MemoryEvery AI session starting from zero is an organizational tax. How source-linked personal memory works today, and where Preview and Planned memory capabilities begin.10 minJun 21“Done” Is a Claim. Evidence Is the Standard.Independent review is available today. This essay separates that current boundary from the planned CI, deployment, telemetry, and evidence-gated completion architecture.9 minJun 14Your Change Does Not Live in One RepositorySeveral agents can edit several repositories. The remaining problem is keeping one contract, landing order, approval boundary, and review record across them.11 min

Open-source Kubernetes agent

Earlier entries about the separate Apache-2.0 Kubernetes operations project, not the harness.

2026

9 entries
Mar 15Kubernetes observability stops before the fixObservability surfaces signals. Operations still require a bounded decision, an approval gate, typed execution, and verification.8 minMar 04Why Approval Gates Must Be Architectural in DevOps AI AgentsApproval gates in DevOps AI agents cannot be UI toggles or confirmation prompts. They must be enforced at the execution engine level, below the model.10 minMar 01AI for CI/CD Pipeline Debugging with Jenkins and SkyfloHow Skyflo's Jenkins MCP tools work; from natural language build triggering to log analysis, parameter-aware job management, and cross-tool debugging.8 minFeb 28Reducing Kubernetes MTTR with an approval-gated AI agentHow an approval-gated Kubernetes agent can gather evidence, propose a bounded change, execute through typed tools, and verify the result.8 minFeb 27Why Human-in-the-Loop Is Non-Negotiable for Production AIReal failure scenarios, architectural safety gates, and why the approval layer must live in the engine; not the UI. A safety philosophy for AI infrastructure agents.8 minFeb 24Agentic AI vs Script Automation in DevOpsBash scripts are brittle. Raw AI is dangerous. Agentic AI with typed tools, planning, and verification gives you the advantages of both without hiding the control boundary.9 minFeb 21An illustrative Skyflo walkthrough: investigating payment-service latencyA representative walkthrough of Plan, Approve, Execute, and Verify. The values and timings are examples, not measured customer results.10 minFeb 17How Skyflo Works Under the Hood: A Unified Agent Architecture for KubernetesA deep dive into LangGraph orchestration, MCP tool protocol design, and why typed tool execution prevents prompt injection in production infrastructure.12 minFeb 13The operational work coding agents still leave behindCoding agents can produce a change. Deployment, incident response, operational verification, and durable context still require a separate control loop.9 min

2025

20 entries
Dec 14Token + Latency Analytics: Building a Dashboard That Engineers Actually UseTurning TTFT/TTR and cost into trends, budgets, and actionable insights across your conversations.9 minDec 07Slack as an Ops Console: Bringing Human‑in‑the‑Loop to Where Work HappensA single-tenant Slack bridge plan: streamed updates, approvals in-thread, and guardrails that don’t feel heavy.12 minNov 30Auto‑Summarization for Long Conversations: Keep Context, Cut the TaxA design for summarizing older turns when you approach context limits-without losing the details operators care about.10 minNov 23Programmatic Tool Calling: When an LLM Should Write Glue CodeLoops, batching, parallelism, and summarization-where code beats prompts, and how to sandbox it safely.13 minNov 16The Case for Tool Search: Shrinking Context Without Losing CapabilityA roadmap post: defer tool schemas until needed, reduce token bloat, and keep the agent accurate under pressure.12 minNov 09Kubernetes Metrics for AI Agents: `kubectl top` Tools and What They UnlockAdding read-only metrics tools so an agent can answer the question everyone asks first: “what’s hot right now?”8 minNov 02Helm Template as a Safety Primitive: Preview Before You Touch the ClusterRendering manifests with inline values, catching surprises early, and building a diff-first culture.7 minOct 26Kubernetes Rollbacks with Confidence: Rollout History + Undo as First‑Class ToolsShipping safe rollback primitives for deployments/daemonsets/statefulsets-and where approvals belong.8 minOct 19Designing a Terminal‑Inspired UI That’s Actually AccessibleFocus, live regions, contrast, and keyboard navigation-what we changed to make a command-center UI work for everyone.9 minOct 12Real-time token metrics: TTFT, TTR, cached tokens, and costOperators don’t trust black boxes. Here’s how we expose LLM latency and usage without spamming the UI.11 minOct 05FastMCP Streamable HTTP: Migrating Off Legacy SSE TransportWhy we moved, what broke, and how Streamable HTTP made MCP communication simpler and more reliable.10 minSep 28v0.3.2: Batch Approvals Without Losing Safety (Approve All, Safely)Designing bulk approval controls that respect read-only tools, remain idempotent, and keep the operator in control.9 minSep 21v0.3.1: Chat Queueing + Server‑Side History Search (UX for Real Operators)Why fast history, debounced search, and prompt queueing matter when you’re triaging an incident at 2am.8 minSep 14Storing Integration Credentials the Boring Way: Kubernetes Secrets + ReferencesHow Skyflo stores integration credentials as Kubernetes Secret references without putting secrets in the prompt.9 minSep 07Jenkins in Skyflo: Secure Auth, CSRF, and Parameter‑Aware BuildsA deep dive into the Jenkins toolset, integration-aware discovery, and why builds must be parameter-first.12 minAug 31v0.2.0: The rebuild. From WebSockets to SSE and a simpler agent coreWhat changed when Skyflo replaced WebSockets with SSE and simplified the agent core.7 minAug 24SSE Done Right: Streaming Tokens + Tool Events Without Melting Your ProxyA hands-on guide to reliable server-sent events for long-running infra tasks, including NGINX hardening.8 minAug 17MCP in Practice: Standardizing DevOps Tools So AI Can’t Go RogueWhy Skyflo’s MCP server exists, how tools are validated, and what “readOnlyHint” really buys you in prod.11 minAug 10Inside Skyflo’s LangGraph Workflow: Plan → Approve → Execute → VerifyHow Skyflo compiles a compact graph, streams progress, and decides when to continue, stop, or request approval.10 minAug 03Why Human-in-the-Loop Is Non‑Negotiable for AI in Production OpsA practical look at approvals, safety gates, and why “agent autonomy” should still ship with guardrails.9 min

The Kubernetes agent is a separate Apache-2.0 open-source project.

RSS feed