2.3 · How it works· 3 min read
What every agent needs before it acts
Before any agent touches a customer, it needs five things: a written job, shared memory, the ability to act inside your tools, limits and a log, and a way to learn. Miss one and the rest don't hold.
A clever agent isn't enough. Two clever agents can still disagree about the same customer, or send them two things that don't match.
Before any agent touches a customer, it needs five things.
1. A written job
Not "find referrals." Something like: "When a customer reaches first value, draft an introduction ask for the person who knows them, within the week."
A written job covers four things:
- Purpose. The outcome it exists for.
- Scope. Which data it can read, which customers it serves, which stages it touches.
- Limits. What it must never do. How often it may contact a customer. How sure it must be before acting.
- How we'll know it worked. Not "asks drafted" but "introductions that happened."
Without it: two agents chase different goals. One thinks a customer is happy because they log in every day. Another thinks they're at risk because they opened a support ticket. The fix isn't a smarter agent. It's one shared definition, written down.
2. Shared memory and handoff rules
Every agent reads from the same record of the customer, and that record stays current.
When a customer moves from one stage to the next, it's written down what goes with them: the promises made, the worries raised, the people involved.
Without it: each agent keeps its own version of the customer, and the context dies at every handoff.
3. The ability to act, inside your tools
An agent that only produces insight is just a report. It needs to work where your team works: update the CRM, draft the email, create the task, tell the right person.
It also needs checks before it acts, so one bad record doesn't become a wrong message to every customer you have. And it needs a plan for when something breaks: log it, try again, tell a person.
Without it: you get an accurate warning that an account is at risk, and nobody does anything about it.
4. Limits and a log
How far it may go on its own. When it must stop and hand over to a person. A record of everything it did, what it saw at the time, and why it chose what it chose.
Without it: when something goes wrong, nobody can tell why, so nobody can fix it.
5. A way to learn
Results flow back in. Which moments led to an introduction? Which alerts did people ignore? Which drafts did people rewrite before sending? In the Sprint, we review this with you every week.
Without it: the same mistakes repeat, and what worked last quarter quietly stops working.
All five, or it doesn't hold
Each one needs the others.
- A written job with no shared memory: agents work at cross purposes.
- Shared memory with no ability to act: everyone agrees, and nothing happens.
- Action with no limits or log: things happen, and nobody knows why.
- A log with no learning: you can see the mistake, and it happens again.
- Learning with no written job: it gets better at the wrong thing.
The thread through all five
What ties them together is context: the right information, in the right shape, at the right moment.
Same customer, same three moments. Whether that turns into one well-timed message or three clashing ones depends on whether anyone designed how context moves.
Questions this chapter answers
- What does an AI agent need before it contacts a customer?
- Five things: a written job (purpose, scope, limits, and how we'll know it worked), shared memory with rules for every handoff, the ability to act inside your tools with checks, limits and a log, and a way to learn from results. Miss one and the rest don't hold.
- Why do our agents disagree about the same customer?
- Usually because they use different definitions. One reads daily logins as happy, another reads a support ticket as at risk. The fix isn't a smarter agent; it's one shared definition, written down, and one shared record of the customer that every agent reads from.