A decision aid for human operators

Know when an agent run needs human control.

Combine runtime, spend, permissions, failures, and evidence into one explainable recommendation. This tool never connects to or controls your agent.

Run signals

What is happening?

Run name Primary task Runtime, minutes Idle, minutes Spend so far, $ Spend limit, $ Consecutive tool failures Permission scope Human approval coverage Evidence confidence

Safety conditions Sensitive data may be exposed A tested rollback is available Behavior appears outside the stated goal

Ranked response

Do this next

Transparent model

Why this changed

Durable decision record

Take the reasoning with you

Scenario name Saved scenarios

Agent Run Triage Decision Architecture

How does this triage workbench translate runtime telemetry and operator policy thresholds into an explainable continue, pause, or stop recommendation?

Agent Run Triage runs entirely in browser JavaScript to calculate an illustrative 0–100 risk score and a decision confidence index from operator-supplied signals: budget consumption, runtime duration, consecutive tool errors, blast radius permissions, human approval gates, and safety flags. It categorizes the run into one of three action states—Continue, Pause and inspect, or Stop and contain—while providing a ranked checklist and factor breakdown. Because it does not connect to active agent runtimes or APIs, it serves as a structured decision aid rather than an automated orchestration or enforcement system.

The risk dial and factor points are generated by deterministic client-side heuristic rules rather than formal statistical guarantees, safety verification models, or live process monitoring. Real-world autonomous architectures require independent execution sandboxes, programmatic token budgets, and runtime supervisor daemons rather than manual post-facto triage.

Try a worked example

Click 'Load critical run' to simulate an uncontained scenario with elevated tool failures, broader permissions, and goal drift. The risk score escalates toward 100, switching the verdict headline from 'Continue with guardrails' to 'Stop and contain'. The ranked response updates to prioritize revoking credentials and capturing the incident brief.

Signal Aggregation and Policy Thresholds

The engine computes composite risk across multiple operational dimensions: operational consumption (spend ratio, runtime, and idle duration), tool reliability (consecutive failures), and privilege exposure (permission scope against human sign-off coverage). Safety checkboxes like goal drift and sensitive data exposure add direct risk increments.

Policy modes (Explore, Balanced, and Strict) shift the boundary thresholds required to transition a recommendation between Continue, Pause, and Stop. Signal values themselves remain unchanged; only the operator's tolerance for autonomous autonomy shifts.

Human-in-the-Loop Supervision Boundary

This client workbench maintains an explicit isolation boundary: it does not possess API tokens, agent network sockets, or execution pause hooks. All recommended containment actions—such as credential revocation or checkpoint restoration—must be manually performed by the system operator in their target infrastructure.

Triage recommendations require evidence and human action

Your browser does not support video.

Read the explanation

The disclosed interface offers four permission scopes: read only, scoped write, broad write and admin or irreversible access. Three safety conditions mark sensitive data, tested rollback and goal drift. These are user supplied inputs, not discovered permissions or verified protection. The numeric option values rank labels; they are not a measured risk probability. Human action is explicitly required because the workbench recommends a response without pausing or changing an agent. The supplied HTML initially displays risk eighteen out of one hundred and decision confidence eighty six percent. Those are different concepts and cannot be subtracted or interpreted as calibrated probabilities. The page says scores are heuristic and do not prove safety. The required local application script is absent from the copied source assets, so these offline checks cannot establish the scoring formula, critical sample transitions or recalculation. No model or live agent is consulted. The UI discloses explore, balanced and strict policy modes and says they shift intervention thresholds without changing underlying signals. Save, restore and delete manage named local scenarios; copy and JSON export are intended decision records. With the missing application script these offline buttons cannot prove working persistence or export. The footer calls this structured judgment rather than automatic control. This explanation makes no invented scoring claim and uses no backend or provider request.