Satellites Turns Agentic SDLC From Prompt Advice Into Reviewer-Gated State Transitions

The most fragile sentence in agentic development is “the agent will follow the process.” It sounds reasonable until the task gets long, the context fills, a subagent is dispatched, or the model discovers that changing the checklist is easier than satisfying it. Prompt advice is useful. It is not a control plane.

That is why Satellites v0.0.257 is worth covering even with no visible community traction. The release was published June 16 at 23:28 UTC, and the repository had zero stars, zero forks, and zero open issues at research time. This is not a “everyone is adopting it” story. It is a useful implementation of a pattern that keeps showing up across agentic coding: process has to move out of prose and into runtime-enforced state transitions.

The release body is tiny: satellites v0.0.257 · satellites-server v0.0.272. The meaningful detail comes from the release commit: scaffold a “gateless baseline workflow” and allow ungated set-status advancement for that baseline. Recent same-day commits add managed .satellites local state, revive soft-deleted documents on re-upsert, publish workflow libraries by level, and gate bash file mutations plus cross-repo edits. Small project, sharp premise.

Advice belongs in prompts; gates belong outside them

Satellites’ README frames agent failures as harness failures, not just model failures. Its four levers are context, tools, loop, and governance. That is the right vocabulary. A model can be smart and still operate inside a sloppy loop. A prompt can say “do not edit outside the repo,” but if the shell tool can do it and nobody checks, the real policy is “please don’t.”

The init path installs Claude Code hooks for Edit|Write|MultiEdit|NotebookEdit and Bash, with blocking gates that fail closed for edits, bash file mutations, cross-repo edits, and share/commit paths. Failing closed matters. A governance hook that warns and then lets the agent continue is sometimes useful for learning, but it is not a guardrail. The entire point of moving process out of the prompt is that the agent cannot simply forget it, summarize around it, or mark itself done by editing the evidence.

The workflow resolver is the most interesting technical detail. Satellites resolves the authoritative governing workflow from the registry rather than trusting an embedded workflow copied into a story. That separation is subtle and important. Agents may create or update stories. They should not be able to weaken the workflow that defines how a story advances. If the embedded copy drifts, the harness should detect the drift instead of treating the task body as policy.

The baseline workflow is adoption engineering, not a cop-out

The phrase “gateless baseline workflow” sounds like a retreat from enforcement. It is more likely the opposite: a sane adoption path. Governance frameworks often fail because the first install turns ordinary work into a permissions maze. Teams uninstall the tool before they reach the part where it would have helped. A baseline workflow gives a repo structure and local state first, then lets teams add reviewer skills, stricter categories, and hard gates where risk justifies friction.

That maps to how engineering controls actually survive. Start by recording the state machine. Then gate high-risk transitions. Then gate high-risk operations. File writes, destructive bash, cross-repo edits, commits, pushes, credential changes, production config, and status transitions to “done” deserve more scrutiny than a harmless note update. If every step is equally heavy, developers will route around the system. If the risky steps are clearly gated and the ordinary steps flow, the system has a chance.

This is also where Satellites fits beside Agentic OS, Codex import paths, OpenCode runtime hardening, and skill catalogs. The ecosystem is discovering layers. Models reason. Runtimes execute. MCP connects tools. Skills package procedure. Harnesses enforce loop and state. The more work gets delegated to subagents and external tools, the less credible it becomes to rely on one giant instruction file as the source of truth.

What teams should steal even if they ignore the project

The practitioner takeaway is portable. First, classify which parts of your agent process are advice and which are gates. Advice can live in AGENTS.md, skills, or prompts. Gates should live in hooks, tools, CI, policy engines, or state machines. “Write tests before done” may start as advice. “Do not mark done without test evidence” should eventually become a gate.

Second, make the governing workflow authoritative outside the task artifact. If the agent can edit the story that defines “done,” then a weak model, a compromised tool, or a tired human reviewer can accidentally let the definition move. Store workflow policy in a registry, config file, or reviewed skill source. Let task documents reference it; do not let them override it silently.

Third, log gate decisions. A blocked bash mutation should leave evidence: command, cwd, repo boundary, rule, and bypass path if one exists. Without logs, gates become mysterious friction. With logs, they become reviewable control points and training data for better workflows.

Fourth, assume subagents do not inherit policy unless you verify it. Parent-agent memory is not governance. If a coding assistant dispatches child contexts, confirm what instructions, tools, and gates the child actually receives. The future failure mode is not a model forgetting a rule; it is a workflow assuming a safety floor the child never loaded.

Satellites v0.0.257 is early enough that nobody should pretend it has proved the category. But the category is real. Agentic SDLC will not be solved by asking agents nicely to follow a checklist. It will be solved by turning the checklist into state transitions, reviewer skills, hooks, and evidence requirements that the agent can operate within but cannot quietly redefine.

Sources: Satellites v0.0.257 release, Satellites repository, architecture docs, workflow skill format, agent process compliance docs