Solve it once. Run it forever.
Merlin missions are TypeScript files that encode your expertise — which model runs each step, where you must approve, what always passes. Commit them once. They run for every engineer, on any device, autonomously until a checkpoint needs you.
“What if knowing when to intervene was something you could ship?”
Missions
Not just guidance for an agent to interpret. A mission is executable control over how the work runs: which model handles each step, how contexts split and pass state, where humans interact, what always gets validated, when the system must stop, and what logic the workflow can express.
Without Merlin
The agent runs. No one decided where to check in. Step 12 drifts. Steps 13–20 compound the error. You find out at the end — if you find it. And nobody on the team makes the same call twice.With Merlin
You describe the rules: research runs freely, plan needs approval, execution stays scoped, tests are non-negotiable. Merlin writes that as a mission. It enforces your rules for every engineer, every run, without you watching.A mission is a TypeScript file that encodes how a task should be done — not just what to do.
It’s also what you fire off when you need to focus on something else. The mission runs, validates, and halts at the exact moments that need your judgment. Ten missions across ten projects — you’re notified when any one of them needs you.
Your staff engineer describes the workflow intent; Merlin generates the mission and commits it to .merlin/missions/. It gets code-reviewed like any other code. Versioned. Shared across the team.
Architect
Senior / Staff Engineer
.merlin/missions/Any Engineer
Junior, Mid, Senior — Everyone Else
Mission vs Skill
A skill is Markdown: quick to write, easy to tweak, weak on control. A mission is the inverse: more deliberate to author, but backed by full TypeScript and exact execution rules.
Easy authoring, weak control
Skill
More effort, explicit control
Mission
A skill tells one agent something useful. A mission tells Merlin exactly how the whole job must run.
Skills optimize for ease. Missions optimize for control and reliability.
Pick a mission. Provide inputs. Wait for results. Approve the plan when it appears. Decide whether to commit when it’s done. No AI knowledge required.
Monitor execution in real time. Make decisions at branch points. Override model selection per step. Inspect every agent decision before it becomes code.
Design the workflow: research strategies, validation criteria, approval checkpoints, model per step. Ask Merlin to generate the TypeScript. Review it, optionally edit it, commit it.
Most of your team runs at Level 1–2.
Your top engineers work at Level 3.
Everyone ships Level 3 quality.
Mission Code
A mission is the same workflow every time: research first, approve the plan, execute only what was approved, validate the result, then stop for a final human call.
Five checkpoints. Two human approvals. One consistent standard for how work runs.
Research
Read only. Gather context.
Plan + approve
Draft the fix. Wait for approval.
Execute
Ship only the approved plan.
Validate
Run tests. Review the diff.
Commit decision
Human decides what ships.
// Research with read-only tools
const research = await agent("Analyze related patterns in the codebase", {
tools: ["glob", "grep", "read"],
model: "gemini-flash",
provider: "google"
});
// Draft a plan, then stop for approval
const plan = await agent(
"Propose a fix using this research: " + research
);
const approved = await prompt("ApprovalForm", {
title: "Review the proposed plan",
plan: plan,
options: ["Approve", "Approve with changes", "Reject"]
});
// Execute only the approved work
await agent("Implement the approved plan: " + approved);
// Validate with tests and a regression review
await bash("npm run test");
const review = await agent("Review the diff for regressions");
// Require a human decision before commit
await prompt("ResultsView", {
title: "Review changes before committing",
diff: review,
actions: ["Commit", "Request changes", "Discard"]
});Five steps. Two human checkpoints. One mission file that enforces the same engineering standard every run.
Everything your missions need to run, monitor, and improve.
The Switchboard shows every mission across every project, sorted by urgency — input needed, error, running, done. One view, full triage. Live metro map of each run. Get notified when a checkpoint needs you. Nothing to watch until then.
Full audit trailUse GPT-5.4-mini or Gemini 2.5 Flash-Lite for research, GPT-5.4 or Gemini 2.5 Pro for planning, Claude Opus 4.7 where rigor matters, then Qwen3 or Gemma 4 locally for focused execution. Mix Anthropic, OpenAI, Gemini, and local models in one mission.
Per-step budget controlsCan’t skip research. Can’t bypass approval. Can’t commit without review. V8 isolate crash isolation.
Architecture-enforced constraintsFire a mission from your laptop. Approve the plan from your phone on the train. The machine keeps running. Pick up results on any device.
Cross-timezone handoffFor your highest-stakes steps: 3 providers vote, majority wins. The architect places this at the nodes where per-step accuracy matters most.
Multi-provider validationEvery run audited. Engineers rate missions. Teams see what works. Missions improve with use.
Institutional knowledge retentionWhy Now
AI agents reached capability. Orchestration didn’t keep up. 13% fewer junior engineers are entering the field — the apprenticeship model is breaking. “Vibe coding” proved that speed without discipline produces debt, not software. MCP standardized tool integration. The window to encode expertise is now. Sundeep Teki, Stack Overflow 2026Every run that completes sharpens the expertise. Engineers rate missions. Teams see which checkpoint placements hold under real conditions. Missions improve. Your best engineers’ judgment doesn’t retire when they move on — it compounds in git, running for everyone who comes after.
Their expertise — encoded once, running dependably forever.