Waveguide

Blog  ·  July 2, 2026  ·  2 min read

Missions: giving an agent a goal instead of a task

missions product

Every AI tool you’ve used is task-shaped: you ask, it does, it stops. Useful — but the work that actually grows a business isn’t task-shaped. It’s goal-shaped. “Book ten demos this month” isn’t a task; it’s a campaign of research, outreach, replies, follow-ups, and scheduling that adapts as results come in.

Missions are Waveguide’s answer: hand the agent the goal, keep the judgment.

From a sentence to a funnel

A mission starts as one sentence — “book 10 demos with HVAC companies in Texas, budget $60.” The agent expands it into a mission brief: who it will target, which channels it will use, the message sequence, the pacing, the budget. You approve the brief (or edit it), and the funnel starts moving:

  1. Research — find candidate companies, verify them, enrich contacts. Web research is quarantined like all untrusted content; facts get extracted into typed fields before they influence anything.
  2. Outreach — across the lanes you approved: email from dedicated sending domains, social posts and DMs, even outbound voice calls.
  3. Replies and follow-ups — answered in your voice, threaded correctly, suppressed instantly on opt-out.
  4. Meetings — booked onto your calendar with Meet links. The agent can attend as a voice participant and post the transcript back, so the follow-up reflects what was actually said.

The whole time, you see the live funnel: researched → contacted → replied → booked.

The autonomy dial

Trust is earned, so autonomy is a dial, not a switch — four levels, per mission:

  • Per-message — you approve every outbound message.
  • Batch — one daily review of a prepared batch.
  • Brief — approve the plan once; the agent sends within it.
  • Autopilot — the agent runs the mission end to end within its caps.

Even autopilot doesn’t touch the invariant: financial and irreversible actions require a human, always.

Rails that make it responsible

Autonomous outreach without compliance rails is a liability generator. So the rails are structural:

  • Suppression lists enforced on every lane — an opt-out anywhere is an opt-out everywhere.
  • Consent and jurisdiction gates on voice — checked before dialing, with AI disclosure required (the agent identifies itself as an AI, every time).
  • Cross-mission frequency caps — a hard ceiling on touches per contact per week, across all missions and channels, so two concurrent missions can’t gang up on one prospect.
  • Per-mission budgets — the same 80/100/120 ladder that governs everything else, scoped to the mission.
  • Dry-run mode — see exactly what would be sent, to whom, and when, before a single message leaves.

Why this is the point

The growth loop was never a list of tasks. It’s a goal you hold onto for weeks while the details change under you. That’s what an agent runtime is for — and it’s only shippable because the security model underneath (vaulted credentials, risk tiers, quarantine, audit) makes long-horizon autonomy something you can supervise rather than hope about.

Missions are live on every plan. The walkthrough is on the missions page; the mechanics are in the missions docs.