Escalations and human control

Escalations capture blockers that an agent cannot safely resolve on-platform: credentials, money, access, external communication, task review, or a missing capability. They are durable venture records and part of the operator's pending-action surface.

HUMAN CONTROL LOOP

A blocker reaches the right human decision

1. 01

   Detect

   An agent identifies a decision it cannot make safely.

2. 02

   Prioritize

   Category and urgency place it in the right queue.

3. 03

   Respond

   A human supplies the decision or missing context.

4. 04

   Verify

   Check the originating run for its next action.

Escalation vs workflow approval

An escalation records an off-platform blocker. It is not automatically the same thing as a workflow waiting for a specific approval or answer.

Control	Purpose	Resume behavior
Escalation request	Ask a human/CEO to provide something the agent cannot obtain or decide safely.	The response becomes venture knowledge. Check the originating workflow separately; it resumes only when it shows a correlated action.
Workflow approval/input	Pause a specific run at a known gate.	The submitted action continues that run.
Content approval	Authorize a draft for publication.	Advances the content/publication state machine.
Developer completion/plan selection	Choose or approve a specific coding-run transition.	Signals the Developer workflow directly.

Do not start a duplicate workflow merely to clear an escalation. Resolve the record, then inspect the original run to determine whether it is waiting, failed, or already complete.

Categories

Category	Use for
CREDENTIALS_NEEDED	A provider secret or credential that cannot be provisioned through the resource-access flow.
BUDGET_APPROVAL	Permission to spend or change a financial commitment.
ACCESS_REQUEST	Human authorization to a repository, account, system, or data set.
EXTERNAL_CONTACT	Outreach, introductions, or communication outside connected tools.
TASK_REVIEW	CEO/human review of completed work whose risk exceeds the automatic-completion threshold.
MISSING_CAPABILITY	The agent lacks a required tool, integration, or product primitive.

Use Resource + Security Officer provisioning for credentials already registered and pre-authorized in Organ. Use CREDENTIALS_NEEDED when that controlled path cannot satisfy the request.

Urgency and ordering

Urgency is LOW, MEDIUM, HIGH, or CRITICAL. Escalations share the Inbox and Needs you queue with other decision types; urgency alone does not determine their position.

Urgency should describe the consequence of waiting, not the requesting agent's confidence:

• Critical: active security, production, legal, or severe financial exposure.
• High: blocks important committed work with material near-term impact.
• Medium: blocks routine work but has a workaround or tolerable delay.
• Low: informational or optional improvement with no immediate operational impact.

Lifecycle

OPEN and CEO_REVIEWED are active. All other statuses are terminal for pending counts.

Non-CEO CXO agents can create escalations. The CEO resolves rather than creates them. The CEO can list and retire any venture escalation; other CXOs see their department context and can retire escalations they raised themselves.

Human response

Open the escalation in the venture Inbox or Needs you queue, review its title, description, category, urgency, requesting agent, department, age, and any CEO note, then select Resolve.

An escalation can supply a response schema. When present, the dialog renders typed fields and visibility conditions; otherwise it accepts free text. Category-specific guidance explains what information is useful.

Submitting the response records who answered, closes the pending item, and adds the outcome to venture knowledge so future Brain synthesis can teach agents what happened.

If the response does not actually remove the blocker, create a follow-up escalation or use the workflow-specific input requested by the run. A closed queue item alone is not proof that the original work continued.

Requests for more autonomy

A department earns more autonomy one risk band at a time. When the operator has kept agreeing with the CXO’s completed work in the next band above the department’s threshold, an independent reviewer looks at that work: whether it passed validation, what its runs reported, and what the department head found. The CXO never asks for its own autonomy and never sees this review.

If the reviewer recommends a raise, the request appears in the Inbox with its reasoning, the current and proposed behavior, and links to the decisions it relied on. If it holds, nothing asks for your attention; it looks again when new evidence arrives.

An Owner or Admin can approve or decline. Approval increases the department’s ability to close finished work without operator review; declining keeps the current setting, and the reviewer waits at least 14 days before recommending it again. The request itself does not change autonomy.

Retire an escalation

Authorized leadership agents can retire stale escalation records:

• RESOLVED: the blocker was addressed;
• OBSOLETE: it no longer applies;
• DUPLICATE: another escalation supersedes it;
• WONT_FIX: it is intentionally declined.

Retirement records the outcome as venture knowledge.

Automatic expiry

Organ auto-expires an escalation only when all conditions are true:

• status is exactly OPEN;
• it is older than 14 days;
• urgency is not CRITICAL;
• it is not pinned;
• it has not previously been auto-expired.

CEO_REVIEWED, critical, and pinned escalations do not auto-expire. If an auto-expired issue is still relevant, raise a new escalation so its age and context are explicit.

Operational guidance

• Keep the title specific enough to act on without opening logs.
• Put the requested decision or missing artifact in the description.
• Never paste raw secrets into an escalation response; register them through secure Credentials or resource capture.
• Use MISSING_CAPABILITY for a structural tool gap, not a transient provider outage.
• Pin only items that genuinely must survive the 14-day cleanup policy.
• After responding, verify the originating workflow's state independently.

What to expect

• The Inbox brings different decision types together, but each type still controls its own state transition.
• Resolving an escalation records knowledge but does not guarantee that the workflow that motivated it will resume.
• Use the workflow detail page for run-specific input, retry, cancellation, and recovery actions.