How employees do a job
Every employee follows the same method: understand, plan, do, check, report. Each role has a playbook and each kind of job has its own report shape and step budget.
Updated 2026-10-13
The quality bar does not depend on a model remembering a long prompt. The method, the playbooks, the report shapes, the step budgets and the failure handling are plain code with tests, and the engine applies them around every call.
The method
- Understand it. Before the first tool call, decide what “done” looks like: who the result is for, what they will do with it, and the two or three things it must contain. Read any attached file or linked page first.
- Plan it. For anything with more than two or three steps, call
planonce with the steps and success criteria (a few words each). It shows you what the agent is doing and keeps it on track. Skipped for quick lookups. - Do it, the most direct way first. Take the step that gets to the finished thing soonest. When a tool fails, change the approach. After two real attempts at the same thing, stop and tell you what is blocked and the one thing that would unblock it.
- Check it. Before reporting, check the work against the criteria: open the file that was made, run the code that was written, recount the numbers, make sure every claim has a source or is marked as a guess. Fix what is wrong.
- Report it. Lead with the result, then what was made, then what could not be confirmed. A clear “I could not verify X” is worth more than a confident guess.
Kinds of work
The brief is classified into one of seven kinds. The role breaks a tie and covers a vague brief: a Researcher asked to “write up the competitors” is still researching.
| Kind | Triggered by words like | Step budget |
|---|---|---|
| Research | research, map the competitors, market size, pricing of, find out, benchmark, teardown | 22 |
| Build | build, code, script, website, landing page, prototype, repo, pull request, fix, deploy | 22 |
| Data | analyse, MRR, churn, cohort, funnel, forecast, model, spreadsheet, chart | 22 |
| Writing | write, draft, rewrite, blog, newsletter, press release, pitch, memo, brief | 18 |
| Outreach | outreach, cold email, lead list, prospects, follow up, intro emails, SDR | 18 |
| Ops | checklist, weekly update, Slack, Notion, tracker, process, SOP, agenda | 14 |
| Chat | Everything else, and quick questions | 14 |
Role playbooks
Each role adds a short working standard on top of its persona. Here is the idea behind each.
| Role | The standard |
|---|---|
| Researcher | Frame the question by the decision it feeds. Go to primary sources and open the page instead of trusting a snippet. Collect the same fields for every competitor so they compare. Triangulate any number that matters. Separate facts, inferences and gaps. End with the “so what” for your plan. |
| Outreach | A list of 15 well-qualified people beats 100 vague ones. Qualify against the ideal customer, record the signal and its URL, write each email from that signal in under 120 words with one ask, never invent facts or guess an address without saying so, queue sends for approval. |
| Writer | Settle the reader, the one thing they should do, the proof you really have, and your voice. Lead with the reader’s problem. Use only real proof. Offer options when it helps you choose. Reread as the reader before handing it over. |
| Engineer | Read before changing. Work in small steps and run something after each. Verify like a reviewer: tests, the program itself, a screenshot, one failure case. Use a branch, small commits and a pull request that says what changed and how it was tested. Never touch secrets or force-push. |
| Analyst | State what you will measure before computing. Look at the data first: row counts, ranges, nulls, outliers, units. Compute with code and check a second way. Say what the numbers do not show. Give ranges and assumptions for forecasts. Deliver models as real .xlsx with live formulas. |
| Operator | Turn a vague ask into a checklist with an owner, a next action and a date. Check what already exists before creating anything. Keep updates short and specific. Do not post anything public until you have seen it. Keep trackers current rather than creating new ones. |
What the agent may claim
- Only what a tool result shows. If no tool confirmed it, it did not happen.
- A deliverable is a file that was saved and read back. A script that made it, a screenshot taken along the way or a description in chat is not the deliverable.
- A failure is reported as a failure: what did not work and the one thing that would fix it.
These are checked by the server against the tool results, not taken from the agent’s words. See Reports and delivery.