They remember
Each agent keeps a log of what it took part in, weekly summaries and the beliefs it revises each term. It also keeps a short list of key moments, such as whether you keep promises about cover.
How it works
You lead a complete synthetic school from an inbox and a diary. In a 1,000-student school, roughly 1,900 AI agents play the staff, students and families, and a rules engine decides what each decision causes.
Inbox
Week 7 · first morning, 08:00
Parents
A parent asks why her son was moved to a lower maths set.
Staff reports
Science says every class is on track for the end of term.
Pastoral referrals
A concern about a student's attendance and mood, logged on Thursday.
External bodies
Your board wants the budget forecast before its next meeting.
A week in the chair
A sitting has a target of 45 to 75 minutes of play, still to be tested with the first school leaders who play it. A programme seat covers 12 played weeks, about 9 to 15 hours, and the other weeks are fast-forwarded or delegated.
01
The week opens on the inbox and diary as they stand at 08:00, already filtered through the school's channels.
02
You read, rank, reply, delegate or leave things. Every choice is logged with its time.
03
The engine ticks through each day. Play pauses for a knock at the door, and some events keep the clock running while you decide.
04
You talk to anyone you choose and chair the meetings in your diary.
05
You see only what you would plausibly know by then. Play on, fast-forward, branch or stop.
Every action takes time from the pack's table of what each action costs, and the diary fills as you spend it. Items left unread age by rule, so an unanswered complaint can grow into a formal one.
Inbox
Email, notes, referrals and requests from the board or authority
Diary
Meetings, duties, events and deadlines
Floor plan
Every room by period and who is in it, so you can drop into a lesson
Conversation
One person at a time, in text, with voice to follow
Meetings
Senior team, departments, all staff, parents and the board
Dashboards
The school's data as it was entered, faults included
Levers
Staffing, timetable, budget, curriculum, policy, interventions, communication and delegation
Time controls
Pause, play, delegate, fast-forward and branch
Every lever costs staff hours, money or trust. A lever pulled without that cost met, such as a new policy with no training time, fails later by rule.
Your words as moves
Each reply carries the moves the agent read in your words, from a fixed list of 16. Rules decide what each move does to trust, workload and the case at hand.
A move repeated to the same person loses force, so daily praise soon stops buying trust. If you think a line was misread, you can contest it.
Conversation · Head of maths
Tuesday, period 4
Head of maths
We're two teachers down on Thursday and I can't cover Year 11 again.
You
Leave it with me.
Read as promise
Sent for a second reading on a larger model, and flagged for your debrief.
Your words as moves
Each reply carries the moves the agent read in your words, from a fixed list of 16. Rules decide what each move does to trust, workload and the case at hand.
A move repeated to the same person loses force, so daily praise soon stops buying trust. If you think a line was misread, you can contest it.
Conversation · Head of maths
Tuesday, period 4
Head of maths
We're two teachers down on Thursday and I can't cover Year 11 again.
You
Leave it with me.
Read as promise
The engine
The school lives in typed records: students, staff, families, classes, the timetable, the budget, procedure cases and the relationships between people. Only the engine writes to them.
An agent's output is an intent, such as phoning a parent or refusing a cover request. The engine checks it against the agent's remit and resolves it by rule.
Each school day runs eight steps
1
Events fire
Calendar and random
2
Attendance
Who is in today
3
Lessons
Progress and engagement
4
Incidents
Behaviour and referrals
5
Procedures
Steps and timers
6 · model calls
Scenes
Agents talk and act
7 · model calls
Channels
Records and documents
8
Log
The day's events kept
End of each week
Morale and wellbeing drift, the budget posts and a snapshot is saved
End of each term
Tracking and reports, then the debrief
Each year
Exams and results, leavers, the new intake and the staffing round
Rules are plain functions with named parameters. If you dispute a consequence, you can open the rule and the numbers behind it.
Steps set by law or national guidance are fixed in each system's pack. An agent can carry out a step badly, but it cannot skip, add or reorder one.
Interventions move a rule's parameters by amounts taken from published trials. Each effect is held as a range with a central default, and a programme can choose the setting.
Agents keep their memories outside the school's state, so what a member of staff believes can differ from what is true.
What you see
Information reaches you only through channels, and every channel has its own delay and faults. Reading through the gap between the school's true state and what the channels show is part of the job.
| Channel | Carries | Delay | Typical faults |
|---|---|---|---|
| Registers | Daily and period attendance | Same day | Missed marks, wrong codes, late corrections, out-of-date class lists |
| Tracking | Termly attainment estimates | Each reporting window | Optimism, gaps, uneven scales, late entry |
| Behaviour log | Incidents and sanctions | Hours to days | Under-recording that varies by teacher |
| Pastoral referrals | Wellbeing and child protection concerns | Hours to days | Thresholds that vary by person |
| Finance reports | Spend against budget | Monthly | Lag and miscoded lines |
| Statutory returns | Census and other returns | Set dates | Coding errors, such as a miscoded reason for support |
| Staff reports | Department and team status | Weekly or on request | Spin and omission |
| Corridor and staffroom | Mood and rumour, sometimes an early warning | Immediate | Distortion and selective reach |
| Parents | Complaints and praise | Variable | The loudest families are over-represented |
| External bodies | Requests and findings from those the school answers to | Variable | They arrive without context |
Staff decide what to pass on. That depends on how serious the matter is, how far they trust you, what reporting would cost them and their own threshold for escalation.
A leader who punishes bad news hears less of it, and a leader who thanks people for early warnings gets more of them.
Each channel carries four fault settings: delay, error rate, omission rate and bias. You can lower them with data-quality work or new routines, and each costs staff hours.
Faults also drift, so a department under pressure records less. None of the settings comes from real student records.
The cast
A 1,000-student school holds about 79 teachers plus support staff, every student, one agent per family and a handful outside the building, from inspectors to social workers. Agents think only when something reaches them, so running cost grows with conversations and barely moves with the size of the school.
Each agent keeps a log of what it took part in, weekly summaries and the beliefs it revises each term. It also keeps a short list of key moments, such as whether you keep promises about cover.
Candour, conscientiousness, ambition, conflict style and appetite for risk are drawn from ranges in the pack, so each staffroom has its own spread of temperaments.
News travels through cliques. Scenes you never watch still happen and still have consequences, and you can replay any of them in the debrief.
Each prompt holds only what that role could know. A reply that relies on a fact outside the agent's knowledge is checked, regenerated and logged.
Research on language-model agents documents how they fail. Each failure has a counter in the design.
They agree too readily
Goals that conflict with yours, a conflict style in every persona, and an engine that ignores flattery when it resolves intents.
Voices flatten into one
Dispositions and histories drawn by the engine supply the variation, because raising a model's temperature does not.
Simulated students are too able
The engine decides attainment, attendance and behaviour, and the model only voices those decisions.
A trainee talks an agent out of role
Your words reach agents only as speech in the scene, never as instructions, and the engine owns the state.
Agents may never change the school's state directly, invent a statutory or child protection step, claim to be a real person, or produce graphic or sexual content. Every output passes schema and rule checks before it reaches the engine or you.
Save points and branches
Every week ends at a save point. Return to any of them, play the rest differently and compare both runs side by side.
Both runs share the same random draws, so the engine's luck is identical. Agents your decision reaches reply afresh, so part of any difference in how they behave is model chance, and branches carry an illustrative label.
Programmes can switch branching off, so a cohort lives with one run as a real school leader would.
Watch two copies of a school splitA run and one branch
Main run
Branch from the end of week 3
The debrief
A coach leads the debrief, and a facilitator sets up cohorts and storylines; one person can do both. Either sees your record only with your consent.
01
Up to 30 of the people you dealt with most each give a short account in their own voice, limited to what they knew.
02
Every decision, its time and what it touched in the engine. A coach can open the cause chain behind any outcome.
03
The vision you wrote at the start of the year, set beside where your diary hours and the budget went.
04
The engine's measures for the term beside what the channels told you, so you see what they hid.
05
Moments mapped to the leadership standard your system uses, without a score.
06
If you branched, the two runs side by side, with a private question about what you could have done differently.
Seats
What is built
A prototype of the engine already runs a year of attendance and is calibrated in two school systems. Each stage passes a gate before the next begins, on published figures, panels of serving school leaders and the build's own replay and cost tests.
Stage 1 · in build
Gives case files and data courses. Gate 1: both packs meet their calibration targets and every run replays exactly. The identical-policy test and sensitivity analysis apply too.
Stage 2
Gives a school leader a full term to play. Gate 2: serving leaders' predictions, blind transcript panels, identity audits and state checks.
Stage 3
Gives a cohort a term to learn from. Gate 3: a partner cohort plays a term and its debrief.
Stage 4
Built in the order the first buyers ask for.
The results of each gate are published as a short note. See the evidence so far.
Design partners play the first terms and tell us where the school feels wrong. Pilots are free or at cost while Synthetic School is validated.