Track D / 4 volumes

Running a multi-agent team as a founder

A track for founders and leads who are already the operator of more than one agent. The pattern is a small fleet of specialists, a manager that routes work, a session identity, visible spend, and no irreversible action without a named person's approval. You are still the operator. The fleet is not a staff you can ignore.

Sample available Linked samples are ready to read. Other modules show planned outcomes and are marked “Coming soon”.

Volume 1

A small fleet, lightly supervised

Sample lesson available · remaining lessons coming soon

Complete the minimum preflight for each specialist and the manager. Use fictional tasks first; the manager must not inherit specialist access. Session identity is a role description, never an authenticated principal.

The durable idea: specialise the agents, route the work, keep identity and memory short, and stay in the approval loop.

D1.01Read the lesson

Specialists, not one monolith

Completed artifact: Fleet map with two specialist boundaries. Outcome: You can give each agent a clear job, rather than one general assistant that is allowed to do everything.

  • Narrow jobs are easier to supervise, cheaper to run, and easier to revoke.
  • If you cannot say what an agent is for, do not add it.
D1.02Coming soon

A manager that routes

Completed artifact: Routing table showing holds and no inherited access. Outcome: You can put a thin manager in front of the specialists, so work is assigned on purpose.

  • The manager routes and checks; it does not inherit every specialist's permissions.
  • A person should still be able to brief a specialist directly when the route is obvious.
D1.03Coming soon

Session identity and memory

Completed artifact: Short session brief alongside a separate principal/grant map. Outcome: You can start a session with a short identity, a short user brief, and a daily memory, rather than a blank chat.

  • The agent should know who it is, who it works for, and what was already decided.
  • Long, unreviewed memory is how a fleet quietly drifts.
D1.04Coming soon

Model routing by cost and difficulty

Completed artifact: Cost/quality comparison including a deterministic option. Outcome: You can send easy, reversible jobs to a cheaper model and keep a stronger one for judgement.

  • Filing a memory and drafting a commitment are different jobs.
  • Watch spend per job, not only a monthly lump.
D1.05Coming soon

Check-in routines

Completed artifact: Approval queue review with one decision and follow-up. Outcome: You can set a short daily or weekly check-in that surfaces blocks, spend, and exceptions.

  • A fleet with no check-in is a collection of unattended chats.
  • Read exceptions first; do not start with a tour of everything that went fine.
D1.06Coming soon

No unapproved sends

Completed artifact: Approved-envelope card and a rejected changed-attachment case. Outcome: You can keep send, pay, and delete behind a named person's approval, even when the fleet is working for you.

  • "I trust this agent" is not an approval.
  • A draft waiting in a side folder is the normal state, not a failure.
D1.07Coming soon

Cost discipline you can see

Completed artifact: Weekly ledger including human review minutes. Outcome: You can see what a week of the fleet cost, in tokens and in your time, and decide what is worth keeping.

  • Attribute spend to the work it served.
  • A specialist that is rarely right is more expensive than it looks.
D1.08Coming soon

When the stack changes, keep the pattern

Completed artifact: Change checklist with fresh gate, isolation and freeze evidence. Outcome: You can move a fleet to a new tool or model without throwing away identity, memory, gates, or the check-in.

  • Treat a stack change as a release.
  • If the new tool cannot stop or cannot record, it is the wrong shape.

Volume 2

Access, secrets, and a fleet you can defend

Coming soon

The same identity and secret habits Track B teaches, applied to a founder who is both the customer and the operator.

D2.01Coming soon

One identity per agent

Outcome: You can give each agent its own access, so you can review and revoke without shutting the whole fleet.

  • Shared logins make a small fleet look simple, and an incident look huge.
  • One connector identity per agent is a good default.
D2.02Coming soon

The human completes sign-in

Outcome: You can keep passwords and two-factor prompts in your own hands.

  • Agents do not type or copy credentials.
  • A mid-session permission failure is often a tool that updated itself, not a reason to move a secret file.
D2.03Coming soon

Never skip a permission prompt

Outcome: You can refuse flags, modes, or shortcuts that skip a permission or approval prompt.

  • A skipped prompt is an ungated path.
  • If a tool offers you a "just trust this session" switch, treat it as a design smell.
D2.04Coming soon

Scopes that match the job

Outcome: You can grant each specialist only the read and write it needs for its job.

  • A research agent does not need a send scope.
  • Revocation should be something you can do in one sitting.
D2.05Coming soon

Secret scanning the fleet

Outcome: You can run the same commit gate, and the same "do not ingest vaults" rule, across every agent's working files.

  • A specialist that writes notes can leak a credential as easily as a person can.
  • Fail closed when a scan cannot run.
D2.06Coming soon

401 versus 403

Outcome: You can diagnose session and permission failures from error details and grant state, then route the remedy to the authorised person.

  • Inspect the provider error, identity and current grant; a status code alone does not prescribe the remedy.
  • Asking someone to "just re-auth" wastes their time when the grant was never there.
D2.07Coming soon

An audit trail the founder will actually read

Outcome: You can keep a log of what each agent saw, did, and asked you to approve, in a form you will review weekly.

  • If the log is only a vendor dashboard, you will not have it when you need it.
  • Payloads in the log are personal data.
D2.08Coming soon

Red-teaming your own agents

Outcome: You can try to make a specialist over-send, over-read, or skip a gate, and fix what you find.

  • Include one attempt that looks like a normal brief, not only a cartoon jailbreak.
  • Write down what you will not accept before the agent is in a live inbox.

Volume 3

The founder's operating rhythm

Coming soon

A fleet only stays useful if you run it like a small team: short reviews, a weekly write-up you still own, and a way to stop an agent cleanly.

D3.01Coming soon

Daily and weekly check-ins

Outcome: You can run a short daily glance and a slightly longer weekly review without turning either into a ceremony.

  • Daily: blocks, pending approvals, odd spend.
  • Weekly: what to keep, what to change, what to stop.
D3.02Coming soon

What you review: exceptions, spend, blocks

Outcome: You can start every review from the things that did not go as planned.

  • Idle-and-silent is a finding.
  • A week with no exceptions still deserves a look at spend and at any bypassed step.
D3.03Coming soon

On-call for your own agents

Outcome: You can say who is interrupted when a fleet agent is stuck, and how to freeze it.

  • If that person is you, say so, and set hours you will actually keep.
  • A freeze path matters more than a clever retry.
D3.04Coming soon

A weekly write-up the founder still owns

Outcome: You can use the fleet to draft a weekly account of the work, then edit and send it yourself.

  • Organise by goals and results, not by a list of meetings.
  • Close with one reflection and a decision, and check it the following week.
D3.05Coming soon

When specialization costs more than it returns

Outcome: You can retire a specialist that is not earning its keep.

  • Extra agents have a coordination cost.
  • Merge or stop before you add the next one.
  • Count the hidden costs: the exceptions it creates, your review time, and the mistakes you have to catch.
D3.06Coming soon

Documentation for the next session

Outcome: You can leave a short pack a future you — or a colleague — can pick up.

  • Identity, scopes, gates, memory location, and how to stop each agent.
  • If it only lives in a long chat, you do not have a fleet. You have a transcript.
D3.07Coming soon

Adding an agent only when the job is clear

Outcome: You can wait until a repeating job has a name, a source, and a definition of done.

  • A new agent is not a reward for a busy week.
  • Write the job first; create the agent second.
D3.08Coming soon

Stopping an agent cleanly

Outcome: You can revoke access, archive memory, and turn off the schedule without leaving a key behind.

  • Stopping is a designed path, not an improvised deletion.
  • Confirm in the log that the last step was a stop, not a hang.

Volume 4

Keeping a fleet honest over time

Coming soon

How the fleet stays aligned as tools, models, and the founder's week all move.

D4.01Coming soon

Corrections that stick

Outcome: You can turn a correction into a change in the brief, the memory, or the gate.

  • Fixing one draft by hand teaches you, not the fleet, unless you record why.
  • Repeated corrections are a design smell.
D4.02Coming soon

Shared memory without shared privilege

Outcome: You can let specialists read a shared memory or wiki without inheriting each other's write access.

  • The second brain is a source, not a shared admin role.
  • Write access to memory should be as narrow as write access to mail.
D4.03Coming soon

Evaluating the fleet, not only one agent

Outcome: You can review routing quality, handoffs between specialists, and the cost of the whole path.

  • A correct specialist with a bad route is still a failed job.
  • Sample real weeks, not only tidy demos.
  • Look at three levels: did each task go well, did the whole path get faster, and did your week actually improve?
D4.04Coming soon

When a tool vendor moves

Outcome: You can re-check identity, gates, and evals when a platform, model, or connector changes.

  • The pattern stays; the implementation is allowed to move.
  • Do not rebuild the operating idea just because the logo on the harness changed.
D4.05Coming soon

Hiring or briefing a person into the loop

Outcome: You can bring a colleague into approvals or operations without handing them every agent's keys.

  • People join the loop as named approvers or owners, with least privilege.
  • The fleet should still make sense if you are away for a week.
D4.06Coming soon

A small set of standing jobs

Outcome: You can keep a short, honest list of what the fleet is for.

  • Standing jobs should be boring in a good way.
  • A long catalogue is a sign you have stopped supervising.
D4.07Coming soon

Cost and quality in the same review

Outcome: You can look at spend and at quality together, so you do not cut the job that was actually working.

  • Cheap and wrong is not discipline.
  • Attribute tokens, then decide.
D4.08Coming soon

The founder remains the operator

Outcome: You can say, in one sentence, what you still decide and what the fleet may finish alone.

  • Autonomy without an operator is just unattended software.
  • The four-part handoff still applies when the only "customer" is you.