Trust

Trust the boundary. Not the prompt.

Every side effect crosses one door. Every answer becomes durable evidence. Every actor keeps a name.

Admission · payments.transfer decision → evidence
IntentAgent
resolved actor agent:billing-ops acting for · user:dana
toolpayments.transfer
tovendor-311
amount$2,400.00
tieract
crossesone call
The door● live
hardline
no denial
authority
grant active
surface
interactive
asksexact intent
Human decisionresolved
Approval requiredact
Transfer $2,400 to vendor-311?

May invoice · net 30

Approve onceDeny
✓ approved by user:dana14:23:44.115scope · this call
01intent recordedtool + arguments
02paused durablyno executor reached
03human resolvedidentity attached
04authority recheckedgrant still active
05executed onceresult journaled
the Agent cannot approve itself a restart cannot erase the pause the receipt outlives the Run
The trust model

Controls live outside the thing they control.

Models propose. Programs order work. Plugins, Sandboxes, Computers, and connectors execute. None of them owns the identity, authority, admission, or record that governs the call.

The control plane stays outside every execution tier. A powerful runtime can do real work without gaining a path back into the loop that decides what it may do.

system boundarycontrol ≠ execution
control plane
01Identitywho is acting
02Authoritywhat is granted now
03Admissionmay this call execute
04Journalwhat became true
execution
Modelproposes
Pluginin-process WASM
Sandboxsession-scoped
ComputerAgent-bound
the machine is never the brainno tier governs itself
01 · one door

Execution requires proof.

The executor only accepts an admission minted for that call. Policy is structure, not prompt etiquette.

02 · one truth

The journal is the runtime record.

Replay, audit, model attempts, decisions, results, and usage resolve from the same event-sourced truth.

03 · real identity

Agents are principals.

The actor and the caller it represented stay attributable—even when one Agent delegates to another.

Authority at both seams

“May” is answered per operation.

Identity enters the platform without carrying frozen authority. Roles, grants, and scope are loaded when the operation is decided—whether the operation changes configuration or touches the outside world.

principal · user:danaidentity only at the edge
01 · management planeConfigure the system
operationcapability.unbindagent:support
authorityload current snapshotrole + grants
Authorizationservice may mutateaccount floor remains
02 · runtime planeExecute an effect
operationpayments.transferAgent call
admissionevaluate this callfloor + grant + policy
Admissionexecutor may runone call · one proof
authority is read nowproof is scoped to one operation
The decision path

Policy is evaluated where execution begins.

The door does not ask whether the Agent sounds confident. It evaluates named inputs in a fixed order and returns exactly one product answer: allow, pause, or deny.

Admission request · support.issue_refundlive evaluation
call context
actoragent:support
surfaceAgent · interactive
tool tieract
amount$184.00
ordered checks
01Hardline floorno denial
02Current authoritygrant active
03Recorded resolutionnew call
04Invocation policysupervised
05Risk decisionact → ask
one answer
ALLOWmint proofexecutor reachable
PAUSEwrite pending gatehuman reachable
DENYrecord reasonexecutor unreachable
a code-level floor no admin can softenunattended would-pause → deny
01 · read only

Explore without acting.

Read-tier calls can run. Act and Privileged calls are denied.

cautious posture
02 · supervised

Act with a person in the loop.

Reads flow. Higher-risk calls pause unless an applicable approval has already been recorded.

human judgment where it matters
03 · full

Run unattended within the boundary.

Calls may run autonomously; live denials, user pins, and the hardline floor still stand.

authority is never ambient
Durable human judgment

A pause is a state. Not a spinner.

The pending gate records the exact call and releases the Run. A restart can happen. A different pod can take over. The human decision still resolves the same held intent.

Resume re-enters the door: the current authority outranks an older approval. The held call executes once—or stays denied.

pending gate · gate_01J9K6● waiting
01 · sealExact intenttool · args · actor
02 · persistPending gatedurable state
03 · waitHuman decisionno active Run
04 · resumeRe-enter doorauthority rechecked
05 · dispatchExecute onceresult recorded
pod Apauses the callRun released
durable stategate + wakeupshared truth
pod Creceives the resolutionsafe to resume
survives process lossresolves across podsresumes once
Live authority

Stable enough to reproduce. Live enough to revoke.

A conversation keeps the tool presentation it started with. Permission does not freeze with it: grants are read at decision time, and revocation reaches every pod within seconds without session affinity.

authority change · capability:paymentspropagation preview
conversation · frozen presentation
filesknowledgepaymentsmemory
names stay stable in context
14:22:04 · admin decision Unbind payments authority ledger updated
all serving pods
pod A currentpod B currentpod C current
no session affinity required
14:22:08agent:billing → payments.transfersame conversation
current grantDENYbinding no longer active
not a mid-flight cancellationtakes effect on the next call—and again on gate resume
journal · support/refund-184source of truth
14:02:11.482Run startedagent:supportthread · client-acme
14:02:12.190Model attempt failedagent:supportroute A · before commit
14:02:12.904Model attempt completedagent:supportroute B · fallback
14:02:13.907Admission pausedagent:supportsupport.issue_refund · act
14:23:44.115Gate resolveduser:danaallow · this call
14:23:45.002Capability completedagent:supportrefund · success
14:23:45.618Run completedagent:supportusage reconciled
replayauditmetricsbilling reconciliation
One truth

The receipt is part of the runtime.

The journal is not a chat transcript with extra fields. It is the append-only source for conversation replay, every model attempt and route, capability results, autonomy decisions, and usage reconciliation.

A provider failure does not disappear behind the answer that eventually worked. A human decision does not become an editable note. “Why did this happen?” resolves to events, not reconstruction.

principal chain · refund decisiondelegation preserved
humanuser:danainvoking caller
Agentagent:supportdelegates billing work
Agentagent:billingresolved actor
Capabilitypayments.refundadmitted call
actoragent:billing
acting foruser:dana → agent:support
decisionallow · admitted
Execution isolation

Power runs outside the control loop.

Each workload gets an explicit execution boundary. Identity, authority, admission, and the journal remain in the platform control plane—not inside the code or machine doing the work.

01 · plugin

Sandboxed in process

Versioned, permissioned customer logic exposes declared tools through the Capability boundary.

compact logic
02 · Sandbox

Disposable machine boundary

Session-scoped Linux work runs away from the control plane and returns artifacts through the platform.

isolated operating system
03 · Computer

Persistent, Agent-bound machine

The environment can retain state. The Agent runtime and the control loop stay on the platform.

the machine is not the Agent
04 · connector

Credentials behind admission

Connector secrets stay encrypted at rest, are withheld from management responses, and decrypt at use.

external execution · governed call
IdentityAuthorityAdmissionExecutionJournal
For the hard questions

Bring a scenario. Get an invariant.

Trust reviews should interrogate failure paths, not adjectives. These are the questions the architecture is built to answer.

QuestionStructural answerEvidence
Can the Agent talk around policy?

The executor requires a call-bound admission proof.

allow / pause / deny event
Can an approval vanish on restart?

The pending gate and resolution are durable state.

gate id + resolver identity
Can revoked access linger?

Authority is loaded per decision and rechecked on resume.

next-call denial
Can delegation blur who acted?

The resolved actor and caller chain travel together.

principal attribution
Can retries disappear behind success?

Every model attempt and route remains on the trail.

attempt sequence
Can unattended work wait forever?

A would-pause decision becomes deny where no human can answer.

surface-aware reason

Build autonomy you can answer for.

One door · one truth · a name on every action