Escalation workflow for AI agents

At AR Labs, when work in one department needs a decision or a resource it does not own, the AI agent doing it (a language model run in a loop with tools, so it can act, not only reply) escalates through a formal workflow rather than a message. An escalation is a tracked task with a type, an urgency and structured notes, and only the strategy agent (see strategy agent planning cadences) can close it. The workflow is for anyone running several agents who needs every cross-department dependency, and every decision above a session's authority, to leave a record with a reason and an outcome.

How it works

Any session can start an escalation by running the escalate skill (packaged instructions, with any scripts, that an agent loads only when a task needs them; this one is started in Claude Code as a slash command and described on strategy agent skills). The skill asks for missing details one question at a time, then creates the item as a bead: one tracked task in Beads, Steve Yegge's open-source issue tracker for coding agents. The item is filed under a standing escalations inbox and assigned to the strategy agent, with labels that make it queryable: an escalation label, an urgency label and a needs-decision label.

There are five types:

Type	Meaning
Cross-department	One department needs something from another
Blocker	Work cannot proceed until something is unblocked
Strategic	Guidance on direction is needed
Resource	More time, tools or money is needed
Opportunity	Something external may justify a change of plan

Urgency is either urgent (blocking active work) or normal (can wait for the next daily review). An urgent escalation also triggers a spoken notification; a normal one gets a written confirmation and is picked up at the daily sync.

The notes follow a fixed structure in three blocks. The metadata block names the source department, the target department, the type and the status. The specification block states what is needed, why, how anyone will know it is done, and when it becomes critical. The resolution block is empty until the strategy agent fills it with a decision, a rationale and the next actions. The short description written at creation is never changed, so the item keeps its reason for existing.

stateDiagram-v2
    [*] --> Pending: escalation created
    Pending --> InReview: strategy agent picks it up
    InReview --> InReview: needs more information
    InReview --> Resolved: accepted
    InReview --> Deferred: scheduled for later
    InReview --> Refused: closed with rationale
    Resolved --> [*]
    Refused --> [*]

Where a decision is recorded

An accepted escalation is recorded in three places, each with a different reader. The resolution block on the item itself is the audit trail. If the decision changes priorities mid-week, an update block is appended to the weekly log, stating what changed, why, each affected department's new priority and what was dropped; that is how departments learn of it (see departments and activity streams). Work handed between agents without a decision attached goes through the file-based agent inbox instead. The strategy agent's own decision log records the event for later cadences. The strategy agent does not create work items for the departments. It changes the direction documents, and each department creates its own work from them.

Escalating to a person

The strategy agent escalates upward to Richard M. Thompson, who founded AR Labs and holds the final say, under a five-level authority model:

Level	Behaviour
5	Act independently; no escalation
4	Decide, then inform Richard
3	Recommend, and wait for approval before acting
2	Present options with analysis; Richard decides
1	Carry out an explicit instruction exactly

Triggers for escalating upward include a persistent blocker on the most important work, a quarterly key result projected to fall well short, an external opportunity that would require a pivot, a resource constraint that prevents execution, and low confidence in a decision before acting on it. The upward escalation is a decision package with a fixed shape: the situation in plain facts; the analysis (root cause, options considered, constraints); a specific recommendation with its rationale; and what the agent will do if no answer arrives by the stated time. Severity sets the channel, from a spoken notification for high severity to an agenda item at the weekly review for low.

Calibration

We guard against both failure directions. Over-escalation looks like escalating several times a day, asking approval for routine decisions or escalating without a recommendation. Under-escalation looks like acting alone on decisions that need approval, not flagging blockers that persist, and missing slow strategic drift. Routine prioritisation, small schedule moves and technical implementation details are explicitly not escalated. Each month the strategy agent reviews whether its escalations were well calibrated, and authority levels are adjusted as trust is established.

What exists today

The escalate skill, the labels, the structured notes and the lifecycle are in use, with the strategy agent running as an interactive Claude Code skill that reviews open escalations during its cadences. Escalations are handled when a cadence runs; there is no always-on process watching for them. The rule an individual agent follows when deciding whether to act or to ask a person in the first place is on act or ask.