Agency Health Dashboard¶
Metric dashboard — instantiates Agentic Control Loop Design
Turns the live health of an agency loop — is feedback timely, is the actor actually acting, is discretion being used — into a small set of continuously-watched signals.
The Agency Health Dashboard is standing instrumentation that reports, in near-real time, whether an already-built control loop is working — not whether the actor hit its goal, but whether the machinery of agency is intact: is feedback arriving fast enough to update the next move, is the actor exercising its discretion or waiting for permission, are escalations firing at the expected rate, is the model being revised when reality diverges. Its defining move is treating agency as a monitored quantity with leading indicators, so a loop that is quietly degrading — feedback slowing, discretion narrowing, the actor drifting into learned passivity — trips a signal before the missed outcome shows up. It reads the loop continuously; it neither draws its structure nor decides what to do about a red light.
Example¶
An engineering team runs an AI customer-support agent that drafts and sends replies within a set of tools, escalating hard tickets to humans. Outcomes look fine — resolution rates are steady — but the team wants to know whether the loop is healthy, so they stand up an Agency Health Dashboard. Signals: median time from a sent reply to a customer-satisfaction signal (is the feedback arc timely?), the share of tickets the agent resolves itself versus punts to a human (is it using its discretion?), the escalation rate and whether it is climbing, and how often the agent's confidence estimate is contradicted by the outcome (is its model drifting?).
Three weeks in, resolution rate is still flat — but the dashboard shows the self-resolve share sliding from 70% toward 55% while escalations climb. The loop is degrading underneath a stable headline: a recent tool change quietly narrowed what the agent could do, so it is punting more and more to humans. The dashboard did not diagnose the tool change or fix it — it turned an invisible erosion into a watched signal that fell, days before it would have shown up as slower resolutions.
How it works¶
- Instrument the arcs, not the outcome. Each signal measures a link in the loop — feedback latency, discretion used, escalation rate, model-contradiction rate — chosen so that a failing coupling moves the number before the goal is missed.[n1]
- Prefer leading over lagging. A dashboard of only outcome metrics reports the funeral; the useful signals are the upstream ones that predict it, even when the outcome is still green.
- Set bands, not just values. Each signal carries an expected range; the alert is a departure from the band, since "escalations at 4%" means nothing without knowing 1% was normal.
- Stop at the signal. The dashboard raises the flag and hands it off; interpreting why a signal moved, and revising the model, belong to the review ritual, not the display.
Tuning parameters¶
- Signal set — how many arcs are instrumented. More signals catch more failure modes but dilute attention and breed alert fatigue; a tight set is watched but blind to whatever it omitted.
- Refresh latency — real-time versus periodic rollups. Faster surfaces a degrading loop sooner but amplifies noise and can provoke twitchy over-reaction; slower is calmer but lets erosion run.
- Band width — how far a signal must stray before it alerts. Tight bands catch small drifts and cry wolf; wide bands stay quiet and miss the early slide.
- Leading/lagging mix — how much weight sits on predictive arc-signals versus confirmed outcomes. Leaning leading buys warning time at the cost of false alarms; leaning lagging is trustworthy but late.
When it helps, and when it misleads¶
Its strength is early warning on an invisible failure: agency erodes quietly — feedback slows, discretion shrinks, the actor stops trying — and a good dashboard makes that erosion a number that moves while there is still time to act. It also keeps attention on the loop's mechanics rather than only its scoreboard, which is exactly what a stable outcome can hide.
Its failure modes are the classic ones for measurement. Signals get gamed or optimized directly — push the escalation rate down and the loop looks healthier while the agent simply stops escalating things it should. It invites the archetype's own trap: mistaking the dashboard for the agency, treating a wall of green as empowerment when the actor still has no real authority.[n1] And a badly chosen signal set measures what is easy to instrument rather than what actually indicates health. The guarding discipline is to treat every signal as a proxy that can rot, revisit the signal set as the loop changes, and read a green board as "no alarm," never as proof the loop is sound.
How it implements the components¶
The dashboard realizes the archetype's live-monitoring slice — the signals that report loop health, not the structure or the response:
agency_health_signal— it is the set of watched indicators: feedback timeliness, discretion used, escalation rate, model-contradiction rate, each with an expected band and an alert on departure.effect_feedback_loop— it operates the live feedback arc, surfacing the effects of action fast enough and specifically enough that the next choice can react rather than learning only in hindsight.
It reads the loop but does not draw or run it: the static structure of nodes and couplings is Agency Loop Map's observation_channel, selection_policy, and execution_capability, and the world_model whose drift it flags is documented by Model Assumption Register. The map is the one-time diagram; the dashboard is the live readout.
Related¶
- Instantiates: Agentic Control Loop Design — it supplies the live health signal the loop needs to catch its own degradation.
- Consumes: Agency Loop Map tells the dashboard which couplings are worth instrumenting.
- Sibling mechanisms: Agency Loop Map · Model Assumption Register · After-Action Learning Cycle · Briefback or Intent Confirmation · Safe Action Menu · Graduated Autonomy Ramp · Action-Effect Feedback Review
Editorial Notes¶
Form Classification¶
Form family: Monitoring, Sensing & Alerting
Rationale: The mechanism turns the live health of an agency loop — is feedback timely, is the actor actually acting, is discretion being used — into a small set of continuously-watched signals, so its operative form is repeated observation and operational signaling.
Independent corroboration: The frozen evidence defines Agency Health Dashboard as 'Turns the live health of an agency loop — is feedback timely, is the actor actually acting, is discretion being used — into a small set of continuously-watched signals', so its operative form is Monitoring, Sensing & Alerting.
Nearest alternative: Interface, Display & Cue — The display is fed by repeated observation of actual state, so sensing and signaling—not the visual surface alone—is operative.
Review outcome: Independent reviewer agreement; medium confidence.
Origin Attribution¶
Primary origin: Systems Thinking & Cybernetics
Origin pattern: Cross-disciplinary synthesis
Present-day reach: Multi-domain
Rationale: Cybernetics evaluates whether observation, choice, action, feedback, and model update remain coupled tightly enough for an agent to regulate its environment.
Related originating lineages:
- Data Science & Analytics — Operational dashboards turn feedback latency, action rate, escalation, and revision into monitored time-series signals.
- Organizational & Management Science — Empowerment and frontline-autonomy research identify discretion use, escalation, permission delay, and action rights as leading indicators of organizational agency.
- Psychology — Learned helplessness and self-efficacy explain why an actor may stop using formally available discretion after repeated ineffective action.
- Ethics of Technology & AI Governance — Autonomous-agent oversight and detection of shrinking effective authority provide a prominent modern application.
Review resolution: Cybernetics supplies the monitored control-loop object. Operational analytics, organizational discretion, learned helplessness, and autonomous-agent oversight materially form the dashboard's leading indicators.
Attribution caveat: The measured object is a cybernetic agency loop; the health indicators synthesize organizational and psychological agency with analytics.
Encyclopedia synthesis: The exact catalogued form synthesizes established practice rather than reproducing a single standard historical label.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
[n1] A leading indicator moves before the outcome it predicts, unlike a lagging indicator that confirms it after the fact; a health dashboard earns its keep on the former. The archetype's warning that "a dashboard without action rights is not agency" is the standing caution against reading a green board as empowerment. ↩a ↩b