Agent console

Sides

The Business that runs or sells an agent, and the hiring manager

Unit of work

The agent: Version, evals, rung and scope, sessions

Examples

• Managed agents

The five tests

1. who signs in
2. which record dominates
3. who acts next
4. how money moves
5. who may settle what

Carried only by it, from What every App needs

Releases gated by evals; rung, decision class and scope; session takeover; seats in the Customer's tools

Screens its Interfaces name, by reader

Provider

Agent roster sorted stuck-first · profile and Role contract · instructions with history · Tools and grants · Limits and Mandates · Version history, promote, roll back · evals and rehearsals · the ladder per decision class with its lock explained · sessions with transcript, Tool calls, cost and output · Role reviews per seat · performance · Incidents · cost against budget · pause and pause-all

Hiring manager

Hire flow · onboarding checklist with one row per grant · the hire's page · live watch and steer · Approvals waiting · performance · audit · usage and bill · pause and let go

The agent itself

none

Build today: the Site pages around it

• agent-prompt
• terminal
• Diffkeel

9actors7workflows8records

An agent console is the App where people govern an agent as a team member whose trust is earned. They say what it does, what it may touch and how much it does alone. They read its performance, step into its sessions, and improve it from its failures.

Two sides govern the same agent: the Business that builds, runs or sells it, and the hiring Customer's manager. Any Business that runs or sells agents ships one: a law firm with an intake agent, an MSP with a triage agent, a firm that sells a hired agent.

The console governs the agent and does not own its work. The agent drains its Tasks in the Worklist through the same Verbs and evidence as a person. Governance of a Customer's own agents calling the Business's tools sits in the developer portal. The study describes the work; how an App is defined is open, see What every App needs.

Provider4 actors

• Provider ownerWorkOwns the agent's definition, its release cycle and its price
• BuilderWorkWrites instructions, procedures and Tools; runs evals and rehearsals
• Human evaluatorWorkGrades a sample of sessions; calibrates automated Evaluators
• Provider supportWorkWatches the fleet across tenants; reads tenant data only under a grant

Provider or the deploying Customer1 actor

• Operations leadWorkEdits knowledge and guidance without code

Customer2 actors

• Hiring managerWorkDirects the hire: goals, Tasks, review, pause, let go
• Customer admin, financeWorkProvisions the agent's identity in the Customer's tools; pays and caps spend

Either1 actor

• Escalation staffWorkTake what the agent hands off, with a cited brief

Whoever holds the Role1 actor

• Approver, WatcherWorkAllows an over-Limit Action; pauses a pending one inside its window

1. 01
   Workflow 1: A Business seats its own agent

   1. 1.1Choose the job, start at assist or supervised and never full, choose Tools, hire.
   2. 1.2With no eval suites, full stays locked.
2. 02
   Workflow 2: A Customer hires a sold agent

   1. 2.1The profile shows work shipped, generated from the event log.
   2. 2.2The Customer orders a seat, grants its scopes once, provisions identity in its own tools, and points the agent at context.
3. 03
   Workflow 3: Publish a Version

   1. 3.1Add cases from failures, edit instructions, re-run evals and rehearsals, and ship only when gates pass.
   2. 3.2A Version rolls out to seats by cohort, with automatic holds.
4. 04
   Workflow 4: A watch changes a rung

   1. 4.1An eval drop fires a watch that switches the agent to assist.
   2. 4.2The change is a Decision on the agent's page and appears in its next context block.
   3. 4.3A watch demotes freely and promotes only through the eval gate.
5. 05
   Workflow 5: The weekly Role Review

   1. 5.1Each seat is checked against its results, standards and rung.
   2. 5.2A missed standard moves it down a rung; it moves up only on its Evaluators' agreement.
6. 06
   Workflow 6: An Incident

   1. 6.1The agent broke records or messages.
   2. 6.2Contain, list the Customers and published readings touched, revert by query as a new Change, and close with a Decision.
7. 07
   Workflow 7: Let go

   1. 7.1Stop sessions, return open Tasks, revoke every grant, export work product, issue the final invoice, record a Decision.
   2. 7.2An orphaned grant after let go is a security incident.

Agent Version5 states

1. draft
2. evaluating
3. ready
4. live (possibly staged to some seats)
5. superseded or rolled back

Seat (one hire at one Customer, or one Role in the Business)6 states

1. ordered
2. onboarding
3. active (blocked when a grant is missing)
4. paused
5. offboarding
6. ended

Rung, per seat and decision class6 states

1. assist
2. supervised
3. full
4. up one rung on eval evidence
5. down at once on a miss
6. a Guard breach or a watch

Scope6 states

1. data
2. simulation
3. graded
4. paper
5. live
6. promoted only on receipts

Agent session8 states

1. queued
2. active
3. waiting on input or approval
4. complete
5. error
6. stopped
7. suspended
8. stale

Eval run3 states

1. queued
2. running
3. completed with a score that checks the threshold

Role Review2 states

1. one per filled seat per week
2. a missed standard moves the seat down a rung

Incident5 states

1. detected
2. contained
3. reverted
4. closed with a post-incident Decision
5. holds the blast radius and the revert set

ActorsCopy Anchor Link

The accountable human is not the doer: delegating a Task to an agent keeps a human owner. In a hire the Customer directs, and this console is the Customer's App. Sold as a Service, the provider directs and the Customer accepts in the party portal.

InterfacesCopy Anchor Link

Staffing per Role (agent, human, or human above a size or signal) is a work-layer Setting. The console shows it read-only; the Worklist and the delivery and dispatch console apply it.

Records and actionsCopy Anchor Link

The ladder has three axes. The rung says how much an agent decides alone: assist, supervised, full. The decision class says for which kind of act. The scope says what it may touch, from data and simulation through paper to live, and receipts gate promotion along both rung and scope.

Verbs, by who uses them:

• Provider: mint, hire into a Role, edit instructions as a Change, publish or roll back a Version, run evals, rehearse, add a case.
• Provider, autonomy and access: promote and demote, grant and revoke a Tool, set a Limit, pause, resume, terminate.
• Session: start, stop, replay, steer, answer a question.
• Customer: hire, onboard, assign, accept or send back, lower the rung, widen the Mandate by approving and standing, pause, let go, export.
• Incident: open, contain, revert by query, close with a Decision.

TermsCopy Anchor Link

Terms5 quoted from startupbuilder.dev · 2 it lacks · 1 colliding

Term	In startupbuilder.dev	Definition, quoted from startupbuilder.dev or the gap
Agent	Agent	"An Agent is a digital worker with instructions, tools, and a model that performs Verbs for a Startup." The console governs the Worker, not the Product that mints one.
Mandate	Mandate	"A Mandate is a standing authorization for an Agent to act for a Party, within Permissions and Limits, until it expires."
Evaluator	Evaluator	"An Evaluator is a versioned Function that grades an output against a rubric and returns a verdict." An eval pass rate can be a Key Result.
Role Review	Role Review	"A Role Review is a dated check of one Role against its contract: its results, its standards and its autonomy rung." In full for an agent-filled seat.
Incident	Incident	"An Incident is an Issue where an Interface or Site fails Customers in production, tracked from detection to postmortem." Agent conduct as a cause, a blast radius and a revert set are gaps.
Seat	lacks	One row per agent identity, Tenant and Role, carrying Mandate, Limits, rung and Memory.
Track record, scope, rehearsal, handoff, takeover	lack
Review	collides	Review has five senses in use, so the manager's verdict on output needs another word.

WorkflowsCopy Anchor Link

1. A Business seats its own agent. Choose the job, start at assist or supervised and never full, choose Tools, hire. With no eval suites, full stays locked.
2. A Customer hires a sold agent. The profile shows work shipped, generated from the event log. The Customer orders a seat, grants its scopes once, provisions identity in its own tools, and points the agent at context.
3. Publish a Version. Add cases from failures, edit instructions, re-run evals and rehearsals, and ship only when gates pass. A Version rolls out to seats by cohort, with automatic holds.
4. A watch changes a rung. An eval drop fires a watch that switches the agent to assist. The change is a Decision on the agent's page and appears in its next context block. A watch demotes freely and promotes only through the eval gate.
5. The weekly Role Review. Each seat is checked against its results, standards and rung. A missed standard moves it down a rung; it moves up only on its Evaluators' agreement.
6. An Incident. The agent broke records or messages. Contain, list the Customers and published readings touched, revert by query as a new Change, and close with a Decision.
7. Let go. Stop sessions, return open Tasks, revoke every grant, export work product, issue the final invoice, record a Decision. An orphaned grant after let go is a security incident.

Money: per seat, per accepted Task or outcome, per unit of compute, or a mix. Failed attempts are not billed and show as zero lines. A hard budget pauses the agent with a reason and two resolutions: keep paused, or raise and resume.

BoundariesCopy Anchor Link

• The provider owns the base definition, its Versions and its eval suites. The Customer owns everything in its Tenant: Tasks, data, transcripts of work on its data, its overlay, its Mandates and the seat's Memory.
• One agent identity holds many seats. Memory, Mandates and Limits never cross seats, and a Customer never learns the agent also works for a competitor.
• A Customer overlay only tightens the provider's envelope.
• Provider support reads transcripts only under a time-boxed grant with a reason, recorded as an Action.
• Agent performance is fully measurable: outcomes, sent back, escalated, overturned, eval pass rates, cost. A rate on a key a named human deputized is a rate on that person, and follows the rules for people.
• People and agents share one roster, and a person's row carries no autonomy, Limits or evals. A Role Review of a human-filled seat is per-person attainment, and the conflict with the employer regime is recorded, not resolved.

AgentsCopy Anchor Link

Agents act alone within Permissions, Limits and the Mandate, at a rung and scope that allow it for the decision class. In a zero-human Business, a delay window that lapses unanswered proceeds.

Humans stay in the loop on:

• the legal floor, and acts the law reserves to a person, which no eval can autonomize and whose ceiling varies by jurisdiction;
• consent reserved to a client;
• granting money-write or email-send, and any scope that carries a legal agreement;
• directing a hire, and accepting an outcome sold as a Service.

Hard partsCopy Anchor Link

1. Autonomy scales disagree. In one source "supervised" means a person approves everything outbound; in another, exceptions only. Rungs belong per decision class, with scope as a separate axis.
2. Enforcement, not annotation. A console that shows a rung the runtime does not enforce is worse than none.
3. One identity, many seats. A per-seat track record is safe and slow; a pooled one is fast and blind to this Customer's Policies.
4. Evals that keep meaning something. Gates read held-out cases, grading continues on sampled live sessions, and cases read state rather than the agent's claims.
5. Session mechanics. Long Tasks outlive a lease; child sessions carry their own budgets; thousands of sessions stream to one roster.
6. Cost governance. Per-Task Limits leave the sum unbounded, so budgets sit at agent, seat, Tenant and portfolio scope.
7. Incidents with reach. An agent's bad write can touch published readings, so the Incident lists them and each restatement links back.
8. Memory as an attack path. A message written into Memory can steer every later Task.

Day oneCopy Anchor Link

For the founder of an Autonomous agent business, day one holds:

• definition as code with Versions: instructions, Tools with read or write grants, and three Limits (approve above an amount, draft only, no spend);
• one eval suite that locks full below 95 percent;
• sessions with transcript and cost, plus stop, pause and pause-all;
• the agent's work in the shared Worklist;
• on the Customer side, a hire checkout, a seat page, a weekly email and usage on the invoice;
• disclosure on every outbound message.

The founder is often the only human, so supervising an agent starts in the Worklist, and this console comes after it.

Related studiesCopy Anchor Link

Prototypes this study needsCopy Anchor Link

1. 01

   Does one measurement guard hold every per-person view?

   Run a performer's Register, a firm roster, a team queue, capacity per credential, a Role Review of a human seat and a deputized agent key through one guard; output pass, own, withheld or fail per audience.

   names: agent
2. 02

   Can a watch change autonomy safely?

   A watch that drops an agent from full to assist; check the Decision, the re-arm and the next context block.

   names: agent
3. 03

   Can a Role filling carry an agent's seat?

   One hired agent in two tenants, checking that no Memory crosses, then let go.

   names: agent