Setup
AI agents built for the project
One of the five types on this page, or an agent built for your work; with a checking layer and a log.
- Task definition and limits
- A second agent that checks
- Close follow-up at the start
AI agents
We hand repetitive work to an agent, a second agent checks the result, and a person always makes the decision. What each agent does, what it must not touch and when its work counts as done are all written down.
We start with one job; once we see it working, we add the second.
- Project tracking
- Reads the task board and sums up every morning what is late, what is pending and what finishes today.
- Social media
- Prepares a draft for each channel; the publishing decision stays with a person.
- Code agents
- Writes and tests on its own branch; a separate agent reviews.
- Competitor tracking
- Collects, with dates, the changes on competitors’ pages and in AI answers.
- Reporting
- Produces the monthly figures in the same template and checks them against the source.
05.1The work in three parts
Setting up an agent is half the job. The other half is seeing when it gets things wrong and fixing that.
aWhat we do
- Choosing the jobWe find work that takes a few hours a week, can be written as rules, and has a result that can be checked.
- Task definitionFor every agent I write the goal, inputs, outputs, limits and definition of done.
- Checking layerWe set up a second agent that runs on different instructions from the one doing the work; it only writes findings.
- ConnectionsWe connect to sources such as the task board, calendar, analytics tool and repository, with only the access needed.
bWhat we deliver
- A working agentConnected to the tools where the work is done, with its first runs watched together.
- Written rulesWhat it does, what it does not do, when it stops and asks.
- LogFor every run: what the agent read, what it produced, who approved it.
- Handover noteA short guide to turning the agent on and off and changing its rules.
cHow we measure
- Send-back rateHow many jobs the checking agent sends back for fixes; this should fall over time.
- Human correctionsShare of outputs a person changed before approving.
- Time savedTime spent on the same work before and after the agent, from the team’s own records.
- Error logThe wrong output, its cause and the line added to the rules.
05.2How an agent runs a job
Every type has the same four steps. Pick a type, then click the steps: its input, output and limits open up.
Workflow
Agent that does the work, agent that checks, human who decides
-
1 · Start: Every morning
- Input
- Task board, calendar, meeting notes.
- When
- At the start of the working day, at a fixed time.
- Limit
- Does not write to the board; only reads.
-
2 · Agent that does the work: Writes a status summary
- Output
- A short summary under three headings: late, pending, today.
- Source
- Every item comes with a link to the task on the board.
- Limit
- Does not assign work to anyone or change dates.
-
3 · Agent that checks: Reads the summary against the board
- Checks
- Whether links work, whether dates and people match the board.
- If it does not match
- Sends the summary back to the agent that does the work.
- Limit
- Does not rewrite the summary itself.
-
4 · Human who decides: The project owner reads it
- Decision
- Which task moves up, which waits.
- Record
- The decision lands on the board with a date.
- Why a human
- Priority is the job of someone who knows the team and the context of the work.
-
1 · Start: Weekly plan
- Input
- Calendar, announcements, brand voice rules.
- When
- Once a week.
- Limit
- Does not add topics to the calendar by itself.
-
2 · Agent that does the work: Prepares drafts
- Output
- A draft per channel, within the character limit.
- Source
- The page or document behind every claim.
- Limit
- Does not publish or schedule.
-
3 · Agent that checks: Rules and accuracy check
- Checks
- Whether a figure has a source, whether a banned phrase appears, whether text in the image is readable.
- If it does not match
- Sends the draft back with a list of findings.
- Limit
- Does not fix the text itself.
-
4 · Human who decides: Approves and schedules
- Decision
- Which draft, which day, which channel.
- Record
- The approved text and its date are kept.
- Why a human
- The last word spoken in the brand’s name belongs to a person.
-
1 · Start: Task definition
- Input
- A short brief: goal, limits, definition of done.
- When
- Separately for every task.
- Limit
- Nothing outside the brief is taken on.
-
2 · Agent that does the work: Writes the code, runs the tests
- Output
- Branch, commits, test results and a short report.
- Rule
- Does not write to the main branch, does not use commands that erase history.
- Limit
- Asks before adding a new dependency.
-
3 · Agent that checks: A separate agent reviews
- Checks
- Bugs, security, missing tests, work outside the brief.
- If it does not match
- Sends it back with a list of findings.
- Limit
- Does not change the code itself.
-
4 · Human who decides: Merge decision
- Decision
- Merge, hold or drop.
- Record
- Merge note and date.
- Why a human
- Every change that goes live has a person who owns it.
-
1 · Start: Weekly scan
- Input
- Competitor list, tracked pages, question set.
- When
- Once a week, on the same day.
- Limit
- Does not go into pages that need a login.
-
2 · Agent that does the work: Collects the changes
- Output
- A list of changes: what, where, when.
- Source
- Every line carries a link and a date.
- Limit
- Writes no opinions, only the change.
-
3 · Agent that checks: Verifies the source
- Checks
- Whether the link opens, whether the change is really new.
- If it does not match
- Marks the line “could not be verified”.
- Limit
- Does not write a guess as if it were a finding.
-
4 · Human who decides: Decides what to do
- Decision
- Ignore, watch or open a task.
- Record
- The task that was opened, with its source.
- Why a human
- Responding to a competitor is a strategy decision.
-
1 · Start: End of month
- Input
- Connected data sources and last month’s report.
- When
- The first working day of each month.
- Limit
- Does not write to the data sources.
-
2 · Agent that does the work: Produces the report
- Output
- A one-page report and the raw table.
- Rule
- Findings, interpretation and estimates under separate headings.
- Limit
- Does not fill missing data with estimates.
-
3 · Agent that checks: Checks figures against the source
- Checks
- Totals, date ranges, units.
- If it does not match
- Sends it back with the row number.
- Limit
- Does not correct the figure itself.
-
4 · Human who decides: Interpretation and decision
- Decision
- The three jobs for next month.
- Record
- Report and decision in the same place, dated.
- Why a human
- The reason behind a number is known by whoever knows the work.
05.3Related packages
Agents can be set up on their own or as part of another piece of work.
Visibility
Client dashboard
A dashboard showing on one screen what the agents are doing, what they sent back and what is waiting for approval.
- Work waiting for approval
- What was sent back and why
Security
Compliance and risk review
A review of which data the agents can reach, which access they run with and where secrets are kept.
- Access list
- Personal data and secrets check
Continuity
Regular reporting
Monthly reports from the reporting agent, with the figures checked against the source.
Write the job that takes up most of your week.
Who does it, how long it takes, where the result goes; write those three and I will tell you whether it suits an agent.