The Instructions field is where you tell the Agent what to do. It is the setting that decides almost everything about the result.
It also does more than describe the task. Every time you save, Crono reads the description and the instructions to work out which actions this Agent is allowed to use at all, and which of those count as its main actions — the ones it pauses on when it is Supervised. Vague instructions therefore produce three problems at once: the wrong work, the wrong permissions, and pauses in the wrong places.
Write them in this order, using short sentences with one idea each.
Purpose — one line saying what this Agent is for. For example: You qualify inbound contacts for our sales team.
What it works on — the record it receives, or how to find the records. For a trigger-based Agent: Evaluate the contact you receive. For a scheduled one: Work on the contacts created yesterday and owned by me.
Criteria — the rules, written so that someone else would reach the same answer. Name industries, sizes, job titles and statuses. Avoid words like “relevant”, “interesting” or “a good fit” on their own.
Actions — what to do in each case, including the case where the answer is no. Name exactly the list, the field and the value to write.
Stop condition — when the job is done. For example: When the contact has been updated, stop.
Exact names. The Agent cannot see your screen, so “the list I have open” means nothing to it. Write the name of the list, the sequence, the field and the status value.
Create the supporting objects yourself. If the Agent should file contacts into a list, create that list first and give it a name that exists only once — for example ICP 2026 – to review, rather than Test. Then name it in the instructions. Asking the Agent to check whether a list exists and create it if not costs credits on every run and can produce duplicates.
A boundary for research. If the Agent may look things up on the web, say what it is looking for, where to stop, and how many sources are enough. For example: Check the company website and its LinkedIn page. If you cannot find the employee count from those two sources, write “Unknown” and move on.
What to do with missing information. Tell it to skip the record or to write a fallback value, rather than leaving it free to guess.
This is the single most common cause of expensive runs. An Agent that is told to research something, but never told when it has researched enough, will keep going far longer than you expect. Every set of instructions should end with an explicit stop condition — and, if the Agent may search the web, with a limit on how much searching is allowed.
Do not point at your screen. Instead of “Add the contact to the list I have open”, write “Add the contact to the list ICP 2026 – to review”.
Do not restate the trigger as a task. A trigger-based Agent already receives the record; it does not know why it was started. Instead of “Every time a contact is created, check whether it matches our criteria”, write “Evaluate the contact you receive against the criteria below”.
Do not leave the negative case out. If you only describe what to do for a good fit, the Agent has to decide for itself what to do with the rest. Write it: “If the contact does not match, do nothing.”
Do not use open-ended research. Instead of “Find out everything you can about the company”, write “Find the company's industry and employee count from its website”.
Do not say “if necessary”. Say what makes it necessary.
Do not ask for actions this Agent is not built for. An Agent will not invent capabilities it was not given. If the instructions never mention searching the web, it will not search the web; if the agent type is not built to create sequences, asking for one will not work.
Do not pile several jobs into one Agent. An Agent that qualifies contacts and writes sequences and tidies lists is harder to review and harder to stop. Create one Agent per job.
Before — too vague:
“Every time a new contact arrives, check if it's interesting for us and if so do something about it. Search online if needed.”
After — specific enough to act on:
You qualify inbound contacts for our sales team.
Evaluate the contact you receive.
It is a good fit if all of these are true:
- the company is in software, fintech or e-commerce
- the company has more than 50 employees
- the contact's job title contains Head, Director, VP or C-level
Use the information already in Crono. Only if the industry is missing,
check the company website; if you still cannot tell, treat it as not a fit.
If it is a good fit: set the contact's status to Warm and add the contact
to the list "ICP 2026 - to review".
If it is not a good fit: do nothing.
When you have updated the contact, or decided it is not a fit, stop.Both English and Italian work. Write in the language you are most precise in — precision matters far more than the language itself.
Test by hand first. Before you let it run on its own, open a record, start the Agent from the chat and watch one execution end to end.
Read what it actually did. Open Agents, select your Agent, and expand an execution to see the list of actions — or click Read chat in the Activity log. This tells you where it went wrong far better than the final result does.
Diagnose from the action list.
Many actions and high credit use — the stop condition or the research boundary is missing.
It did nothing — the criteria are too strict, or you never described the action to take.
It did the wrong thing — make the branches explicit (“if X … if not X …”).
It asked for confirmation in an unexpected place — the main action it settled on is not the one you meant. Make the outcome clearer in the first line.
Change one thing at a time, save, and test on the same record again. Saving re-reads the text and re-decides the Agent's permissions.
Keep a copy of the version that worked before you rewrite it.
For complete set-ups you can copy and adapt, see Agent instruction examples you can adapt.