Missions
A Mission is a standing responsibility a team of agents owns — an open-ended charter like "no conversation movement goes unrecorded," checked on a schedule and dischargeable however the work demands: workflows, skills, tools, or open-ended agent reasoning. It's the third work mode, alongside structured Workflows.
Vocion supports one Agent framework in three work modes:
| Mode | Shape | Use it for |
|---|---|---|
| Operation | one typed step | classify, extract, draft a field, score |
| Workflow | a known, repeatable sequence | weekly report, sales follow-up, support draft |
| Mission | an open-ended goal a team figures out | "explain why booked calls dropped and recommend what to do" |
These aren't different kinds of agent — they're different amounts of structure for a piece of work. The boundary is who decides the steps: in a workflow the author decides (fixed, deterministic sequence), in a mission the team decides (the lead plans or judges what's needed each time). A mission that keeps producing the same plan is telling you it wants to be a workflow — promote it.
Missions are responsibilities. Workflows are procedures. A mission can invoke workflows; a workflow can borrow an agent for one step.
Standing goals, checked by automations
A mission is not a one-off. It's a charter a team owns continuously — a pure GOAL, containing no logic: no steps, no schedules, no triggers.
# workspace/<org>/missions/follow-up-queue.yaml
goal: No lead or open deal ever goes cold — every thread that needs a next
touch gets a prioritized, context-grounded follow-up drafted for approval.
defaultTeam:
lead: revenue-lead
members: [follow-up-coordinator]
autonomyPolicy: {level: 2}
The cadence lives in an Automation:
# workspace/<org>/automations/follow-up-check.yaml
when: { schedule: "0 15 * * 1-5" }
do: { checkMission: follow-up-queue }
On each tick a check run fires: the lead agent gets the charter + a check instruction, reviews current state, does only what's needed right now (search, run skills, propose CRM actions into the review queue), and reports — or says "nothing needed" and stops. No planner, one lead pass, cheap enough to run hourly. Any mission can still be briefed manually for ad-hoc, planned work.
The litmus test: workflows are deterministic, missions are agent-driven. If you can write the steps down and they're the same every run, it's a Workflow. If discharging it takes judgment — what needs doing this time — it's a mission.
The loop
Hire → Brief → Work → Review → Coach → Learn → Promote. You brief the work in plain language; the team plans it, splits it up, runs it, and pauses for your approval where it matters. Your feedback becomes scoped learnings, and repeatable missions become reusable workflows.
Anatomy of a mission
- Brief + goal — the natural-language assignment and what success looks like.
- Team — a lead agent + specialist members (the lead plans; members work in parallel).
- Plan (task graph) — generated from the brief; each task has an owner agent, a type (analysis / creative / synthesis / artifact / diagnostic / action), and a status.
- Autonomy policy — how much the team can do on its own (see below).
- Artifacts — the deliverables (docs, charts, images, reports) produced via the built-in tools.
- Review + coaching — approve gated steps, give 👍/👎 (becomes scoped learnings), and promote to a workflow.
The autonomy ladder
Autonomy is a property of the mission + action, not the agent. The same agent can draft freely in one mission and be tightly gated in another.
- Draft only — analyze and draft; no external actions.
- Ask before action — prepare actions, but approve before anything is sent/changed.
- Act within rules — take pre-approved actions.
- Manage a goal — own an ongoing goal with budget + escalation.
- Improve itself — propose new playbooks/learnings/workflows (changes still reviewed).
Internal work (analysis, drafting, synthesis) is never auto-gated — it produces drafts, not side-effects. External actions are gated at levels 1–2 and allowed within rules at 3+. A gated task pauses the mission as awaiting_review until you approve.
How it runs (reuses what Vocion already has)
A mission is a team run on the existing agent runtime: the lead plans the task graph, each task is dispatched to its owner agent (subagents / runAgentDeep), request_human_review powers approval gates, artifacts are saved by the built-in tools, learnings capture coaching, and every run is stamped with the active workspace_sha for audit. Durable, multi-day, crash-safe sessions (Temporal) are on the roadmap.
Where it lives
- Templates (starting teams):
workspace/<org>/missions/*.yaml→ applied viaworkspace:apply. - Runtime:
services/MissionService.ts+services/missions/{planner,runtime,autonomy}.ts. - App: Dashboard → Missions (the Mission Room: Brief · Plan · Team · Artifacts · Coaching).
- MCP:
mission_start,mission_get_run,mission_approve, … (see reference/mcp.md).
Promotion
When a mission proves repeatable, Promote to workflow drafts a Workflow from its task graph for you to refine and activate — turning open-ended work into a reliable, scheduled routine.