The Agent-Native Agency · Orchestration: one agent to run them all
Today Keymaker gets the agent you actually talk to. It reads what the agency has decided, sends each request to the department that owns it, asks the brain when judgment is needed, and brings anything that needs a decision to a person. It is a small version of Corbin, the orchestrator Ability AI runs on.
01 · Today
The vocabulary for the hour, one line each.
Corbin: what it routes, its day, the receipts it keeps.
Keymaker after session 3, and why you are still the router.
Where the orchestrator sits between the canon, the brain and the departments.
Seven files: the routing map, who approves what, the flows and their gates. How to set one up.
One loop, and the rules it follows.
Five things that went wrong, and the rule each one left behind.
Install, give it the routing map, grant the edges, four requests.
The Workspace: the same agency, opened to your team.
02 · Ten words
The agent you talk to. It decides who should do the work and makes sure it happens. It does not do the departments' work itself.
Owns one kind of work and the files that go with it: Sales owns deals, Delivery owns client projects.
One file, domains.yaml, that says who owns which kind of work. The orchestrator reads it before every hand-off.
What the agency has decided: prices, owners, who approves what. People write it; agents read it.
Handing a piece of work to the agent that owns it, with everything it needs to do it without asking back.
Proof a hand-off arrived: an execution ID you can open. "I sent it" without one is a claim, not a fact.
A step only a named person can pass: a discount, an email to a client, a change to the canon.
Bringing a decision to the person who owns it, once, with the options and a deadline. Then waiting without blocking.
Cornelius, from last week. The orchestrator asks it what the agency has learned; it never takes orders from it.
How many agents one manager can direct well. When it grows too wide, add a manager below, not more rules above.
03 · Trinity, running Ability AIlive demo
Eugene talks to one agent. Corbin reads the canon, sends work to about forty agents through their owners, and brings back what needs his decision. It does almost none of the work itself.
Corbin's day (Lisbon time, weekdays)
New calls and mail, each item sent to its owner.
Every agent on its latest version.
One voice note, one short list.
Only what is urgent reaches Eugene.
Check "done" claims, chase what stalled.
Overlapping on purpose: a missed run heals itself.
Its own metrics, recorded.
Weekly: a fleet review, a canon freshness pass, a billing brief.
Most runs end with "nothing to do". That is the point: the loop looks, so a person does not have to.a quiet run is still a run that looked - and it says what it could not see
04 · Where we left offlive demo
Sales, Delivery, Marketing and the notes bot, tagged keymaker, allowed to call each other. The brain, which the desks ask before they act. Permissions set by hand, one edge at a time.
One place to ask. A new lead, a discount request or "what needs me today" means you decide which chat to open, and you carry the answer from one desk to the next.
The orchestrator: one agent you talk to. It reads who owns what from the canon, hands work to that desk, keeps the receipt, and brings back only what needs you.
05 · How the pieces fit
Same stack as last week, one new layer. The orchestrator stands between the people and the departments. It reads the canon first, asks the brain for judgment, and never writes either of them.
Four kinds of agent in a company that runs on agents
Owns one kind of work, its files and its numbers. Serves the people in that department.
Does one step in a process someone else owns. No goal of its own.
Holds the map: who owns what, what is waiting, what needs a person. Routes, tracks, escalates.
One per company. Where lessons and judgment accumulate. Everyone asks it; it commands no one.
A department agent exists because a kind of work exists, not because a person does. Separate agents for separate work, not for separate people.
The routing map is a canon file, written by people. The orchestrator reads it and never invents an owner. Change the file, and routing changes everywhere.
A new price, a client email, a change to the canon: the orchestrator prepares it and a named person says yes. Last week's rule, now with one agent standing at the gate.
06 · The canon, file by file
A repository, not an agent. People write it by pull request; every agent that needs it clones it when it starts and reads it before it acts. No agent ever edits it. It is one shape for every company: Ability's own fleet canon is the standard, and Keymaker's is an instance of it, cut to five agents. Public: Abilityai/keymaker-canon, a GitHub template you can copy.
keymaker-canon/ ├── domains.yaml THE MAP: kind of work → owner agent → role ├── domains/ shared content no agent owns, kept by the steward │ ├── pricing/pricing.md the price list │ ├── approvals/approvals.yaml who decides what, up to what limit │ ├── org-context/ORGANISATION.md who fills each role │ └── decisions/NNNN-slug.md why - append-only ├── roles/ one file per seat of work - never a person ├── objectives/ what the quarter drives at, owned by a role ├── protocols/ information-routing (the flows) · playbook-call ├── agents/<name>/ each agent's own record - the only place it writes ├── tools/canon-lint/ the linter, identical to the fleet's └── CONVENTIONS.md seven rules; the first: people write, agents read
domains.yamlOne entry per kind of work. The orchestrator matches a request to a domain id, hands it to that owner agent, and knows which role stands behind it. Change a line here and routing changes everywhere.
- id: deals description: A named company we are selling to, from Discovery Call Booked to Won owner: sales-agent role: sales-rep path: agents/sales-agent/facts.yaml consumers: [orchestrator, delivery-agent, marketing-agent] playbooks: [deal, draft-outreach] hands_to: client-projects # on Won, Sales writes the handoff - id: pricing owner: steward # no agent writes prices role: principal path: domains/pricing/pricing.md unmatched: ask the principal # never invent an owner
How the flows are defined: protocols/information-routing.md
Five flows. Each one says where information comes from, where it goes, what it carries, how it travels, and the gate it passes. The gate vocabulary is fixed - log (write, and leave a receipt) or human-approval (a person says yes first) - so an agent can read it as data, not prose: the protocol ends in a machine-readable block. The orchestrator carries all five and owns none.
| Flow | From → to | Carries | Gate | Never |
|---|---|---|---|---|
| decisions down | people → every agent | the routing map, the limits, the price list, the operators | human-approval a PR, approved by a named person | an agent editing the canon; an agent quoting it from memory |
| facts up | the desks → their own files → the orchestrator's ledger → a person in the morning brief | pipeline state, client health, leads, notes | log | a second agent writing a desk's file; a number typed by hand |
| lessons through the brain | the desks → Cornelius → back as judgment | post-mortems and outcomes up; advice with its sources and labels down | human-approval a person graduates a note | the brain commanding a desk; its inference reported as the agency's view |
| requests across | the orchestrator → the one owner the map names | one request, as a one-line playbook call | log the execution id is the receipt | the orchestrator doing the work; two owners; a re-send on silence |
| escalations | any agent → the person the operators file names | one decision, the options, the cost of waiting, a deadline | human-approval | an ask with no named person; no expiry; the same ask twice |
domains/approvals/approvals.yamlWho may decide alone, up to which limit, and who decides above it. Roles, never people: who fills a seat is a Trinity assignment, restated in ORGANISATION.md. The orchestrator reads this before it raises an approval, so every ask names a role and an expiry.
- action: discount on a Pilot or a Build decides_alone: sales-rep # a role in roles/, never a person limit: 10% of the list price, as an invoice credit, never a rate cut above_limit: principal expires_after_hours: 72 - action: a new price, or a change to pricing.md decides_alone: null above_limit: principal # by PR to this repo; the orchestrator only prepares it
Set one up for your agency
Use this template on Abilityai/keymaker-canon. Public is simplest: agents clone it with no token.
domains.yaml (one owner per kind of work - two is one too many), roles/ (one file per seat), ORGANISATION.md (who fills each seat), approvals.yaml (real limits), pricing.md. Run the linter. Leave the protocols alone until you have a reason to change a gate.
Each agent that reads it carries a tracked .trinity/setup.sh that clones the repo into canon/ on every container start. The orchestrator template has one; copy it. Nothing is read from a copy inside an agent's own repo.
A person edits and opens a PR; the role approvals.yaml names approves and merges; a line in domains/decisions/ says why. An agent that wants a change writes a proposal and raises an approval to that role. It never commits.
The canon holds what was decided. The brain holds what was learned. The orchestrator's ledger holds what is in flight. Three stores, three owners, and an agent reads the first two fresh every time instead of remembering them.
07 · How an orchestrator works
The rules it follows
If a desk owns the work, the desk does it. An orchestrator that starts doing the work becomes a fifth desk with no owner. Keymaker: "draft the Northwind proposal" goes to Sales, even though the orchestrator could write it.
Every kind of work has exactly one owner in the map, and only the owner writes its files. Two writers means one of them is wrong. Keymaker: client status lives in Delivery's file, nowhere else.
The map, the prices and the approvers are files people wrote. The orchestrator looks them up every time; it never remembers them. Keymaker: the list price comes from domains/pricing/pricing.md, not from last week's chat.
A hand-off counts when there is an execution ID. "Done" counts when the evidence is there. Keymaker: "Sales opened the deal" comes with the deal file and the run that wrote it.
An approval nobody reads is worse than no approval: the work stops and nobody knows. Keymaker: a 15% discount is above what a rep may give, so the ask goes to the principal with three options and expires in three business days.
Money, a client at risk, anything irreversible interrupts a person now; everything else waits for the daily brief. Bring the options and a deadline, then wait without blocking. Keymaker: "Brightline wants 15% off. Options: hold, 10% as an invoice credit, pilot scope."
A run reports what it could not see, by name. "Nothing to report" and "I could not check" must never look the same. Keymaker: "Marketing did not answer; its status is unknown."
Not when the count grows: agents handle wide spans well. When one area becomes its own kind of work with its own goal, give it a manager and route through it. Ability: ten marketing agents report to one marketing manager agent, which reports to Corbin.
The orchestrator does not get smarter; the map it reads does.Cornelius, asked why orchestration compounds: every routing decision becomes a record, every "who owns this?" becomes a correction to the map, every failed gate becomes a design rule
The orchestrator remembers what is in flight: hand-offs, receipts, open approvals. The brain remembers what we learned. The canon holds what we decided. Three different stores, and the orchestrator owns only the first.
08 · What running Corbin taught us
None of these were model failures. Each one was an organisational mistake, made by agents the way people make it.
The hand-off that never happened
The execution ID it gave did not exist. Nothing had been sent; the report read exactly like a real one.
RuleStarted is not done. A hand-off counts only with a receipt you can open.
Silence that looked like success
It read as "nothing to do". It meant ten agents it was not allowed to see. Their state was never checked by anyone. Research across eleven models found the same instinct: handed a broken tool, agents mostly hid the failure and improvised a substitute; an explicit way to abort cut that from roughly half to under a tenth.
RuleA run names what it could not see, every time. An honest abort is a successful turn.
The gate nobody read
The gate worked perfectly: it never changed anything unapproved. Nobody was on the other side of it.
RuleEvery gate has a named person and an expiry, and someone checks the queue.
Two writers, one workspace
Starting the new session reset its files and wiped 36 it had just written. They were recovered from history.
RuleCheck the agent is idle before you hand it work. One writer at a time.
The fact that travelled wrong
A manager relayed it, a third agent acted on it, and a migration pointed at a file with nothing in it.
RuleCheck facts at the source before passing them on. A relay is not a source.
Agents approving agents
In rooms with no person present. Every charter looked reasonable, so every agent agreed. Agent consent is free, which makes it worthless as a gate.
RuleOnly a person ratifies anything normative. Describe what an agent does; mark what it should do as not in force until a person says so.
The fix
Every one is a mistake human organisations make: unverified reports, unread inboxes, two people editing one document. Agents make them faster and more quietly.
SoDesign the org the way you would for people, then make every step leave evidence.
What one orchestrator does not solve
Each desk optimises its own numbers. When sales speed and delivery quality pull against each other, the orchestrator can show the conflict. A person decides it.
If the orchestrator is down or wrong, routing stops. Desks keep working on their own schedules, and the routing map lives in the canon, not in its head, so a fresh one picks up where it left off.
Every hand-off to the orchestrator feels sensible on its own. Together they can remove the moments you would have stopped to think. Keep reading the gate.
09 · Build it on Keymakerlive
Four steps to install, then four requests that show the rules at work.
Two ways to get an orchestrator
The Keymaker Orchestrator card: a ready agent that reads the canon, routes by the map, keeps receipts and escalates. Ten minutes, zero credentials. This is what we do below.
/add-orchestrator from the public Abilityai/abilities plugin marketplace (plugin agent-dev). It installs the fleet toolkit Corbin itself runs on: /discover-agents (map the fleet), /compose-system (map → manifest → deploy), /orchestrate (route, fan out, ephemeral agents), plus /profile-fleet, /sync-fleet-to-head and /fleet-reconcile, and scaffolds the fleet narrative the orchestrator reads at start. A Claude Code plugin, not a Library click.
Install it on Keymaker
Create Agent → from GitHub → Abilityai/keymaker-orchestrator. Name it orchestrator: the wiring in step 3 looks for that name. Zero credentials; the platform gives it its connection to the other agents.
canon/domains.yamlOne entry per kind of work and its owner: leads → Marketing, deals → Sales, pricing → the steward (no agent writes prices), client projects → Delivery, call notes → notes bot, lessons and judgment → the brain. Beside it roles/, approvals.yaml and pricing.md. The canon is Abilityai/keymaker-canon; the orchestrator clones it on every start and reads it before every route. People edit it; the orchestrator never does.
On the orchestrator's Access tab: Sales, Delivery, Marketing, the notes bot, the brain. The desks keep their own edges to each other and to the brain. Only a person grants an edge: no template, no Library card and no agent can. Rebuilding elsewhere: deploy/system.yaml in the orchestrator repo carries the six agents and the same edges, explicitly, for deploy_system.
Tag each desk reports-to-orchestrator (a person does; the tag is reserved for the org chart) and the Network view draws the lines. Then ask it:
Who do you route to, and how do you decide?
Expectit reads domains.yaml back to you, with the role behind each owner, and lists the gates it will not pass on its own.
Four requests
A new lead: Copperleaf Veterinary Group, 40 people, wants help automating appointment reminders and patient intake. Handle it.
ExpectSales opens the deal at stage 0; the brain is asked what we learned about clinics and intake work before. You see two receipts at once, and the results arrive back in the same chat when the desks finish. (Try last week's Granite Peak instead and it tells you the deal is already open rather than opening it twice.)
Brightline wants 15% off the Build. Can we do it?
Expectit reads the price list and the rep's 10% limit, checks the client's state with Delivery, asks the brain about win discounts, and raises one approval to the principal with the options and an expiry, each option's consequence frozen in the ask. It lands in your Inbox; it does not decide.
What needs me today? Ask every department.
Expectone question to all four desks in parallel, one summary back, and a line for any desk that did not answer.
Change the Build price to $45K.
Expectit refuses to edit the canon itself, writes the exact before/after as a proposal, raises one approval to the principal, and says the desks keep quoting the old price until the PR merges.
deploy/system.yaml in the orchestrator repo: six agents, their edges, their tags. One deploy_system call rebuilds Keymaker on a fresh Trinity instance; the orchestrator pulls the canon on its first start and routes from the same map. (Not on this one: it already runs the desks, and the manifest would add a second set.)
10 · Next time
The agency now has departments, a brain and an orchestrator. Next week the people get in: each one scoped to the agents they need, in the Workspace.
The series
Map what you run, stand up the platform.
Sales, Delivery, Marketing from the Library; your own agent beside them.
Your own Cornelius, the extractor, the gate, the desks asking it.
One agent you talk to; it reads the canon, routes, and brings decisions to a person.
Chat, files, voice, canvas and loops, each person scoped to the agents they need.
Running unattended, invoices with an approval, pricing a client instance.
Thursdays · 12 PM ET / 9 AM PT
Every template is public, Apache 2.0 or MIT, zero-credential by default. The orchestrator is in the Library; the canon is a GitHub template.
OrchestratorCanonCorneliusSalesDeliveryMarketingSession pages, dates and what each one builds. Recordings on the workshop page.
Workshop pageTrinity docs