From there it decides what to do next after every reply, on whichever channel is most likely to get an answer, until the objective is met or a guardrail stops it.
Two kinds
Every mission is one of two kinds, and the board labels each row so you can tell at a glance whether it will finish.Campaign
Works a list of people and finishes. “Win back the 412 customers who have not been in for 14 months.”
Trigger
Stays on and opens one pursuit each time its event fires. “Whenever a form is submitted, reach that lead.”
The loop
Everything a mission does is one loop, run per person until it wins or a guardrail stops it. The names below are exactly the ones on the timeline, so what is described here is what gets recorded.Not a drip sequence
A drip sends step three on day four whether or not anyone replied. A mission re-decides every time something happens.- It reasons between touches. She replies to texts and never picks up, so the next touch is a text. She has gone quiet through six touches, so it stops rather than becoming the reason she opts out.
- It remembers across conversations. What a customer said in March is context for the call in July. Memory belongs to the person and outlives the mission that learned it.
- It ends in an outcome, not in a step count.
Watching one run
Open a mission and you get the people in it. Open a person and you get their thread, band by band, in the order it happened. Here is a real sequence, shortened.
That day-2 line is the whole idea in one row. You wrote a plan, she did something that made the plan wrong, and it departed from the plan and said why.
Guardrails are a ceiling, not a suggestion
Guardrails are copied onto every person the moment their pursuit opens, and the planner is structurally unable to exceed them.
Copying rather than referencing matters: editing a mission tomorrow cannot retroactively loosen what was already agreed for someone being worked today.
Whether to make contact, on which channel, and when are decided by a policy layer, not by a language model, and every decision is written down with its reason. The model words the message once the policy layer has already decided to send one.This is why “why did you contact my customer at 9pm” is answerable from a rule rather than from a sampled token. In lending and healthcare, buyers audit exactly this.
Outcomes, and disputing one
A mission ends in an outcome, and every outcome carries the interactions that produced it, so it can be opened to see the calls and messages behind it. If one is wrong, dispute it from the mission and say why. The charge is held the moment you do. Disputing is limited to owners and admins, because it moves money.Long horizon is not long running
A 45-day loan origination is long-horizon because it spans dozens of decision points, not because a process runs for 45 days. A mission is a durable record; the planner that advances it is a short task woken by a signal or a timer. Cost scales with the number of decisions, not with elapsed time.Getting a good one
- Say the outcome, not the activity. “Book a service visit” beats “call these people.” The objective is what tells it when to stop.
- Set the touch ceiling deliberately. It is the single control that decides whether a mission feels persistent or feels like harassment.
- Start with one list. Watch a few threads end to end before pointing it at thousands.
- Read the reasoning bands. They are where you find what it believes about your customers, and where a wrong assumption shows up early.
What’s next
Coworkers
Who works your missions, and how you hire them.
The Desk
Where a lead becomes a deal you can move and close.