Agent Ops
What an agent did, what it cost, where it failed and where a human decides.
What is this agent allowed to use, and where does a person step in?
A canvas for designing business agents: triggers, agents, routers, guardrails, human approvals and actions as the flow, and under each agent its equipment — model, skills, MCP servers with per-tool switches and ask-first, knowledge and memory — on square ports with dashed links. A searchable palette (drag, or click to equip the selected agent), an inspector form per node kind, design checks that point at the node they are about, a simulated test run that pauses at approvals, undo/redo and auto-layout. Three sample designs: support resolution, invoice processing, lead qualification. Plain data in, no runtime.
Who did what?
An append-only audit table whose actor column tells three authors apart by glyph — copilot, delegated agent, person — and whose result column is a status badge with an icon (Applied / Held / Stopped / Reverted), so a row still reads in greyscale. Monospace tabular timestamps; the action is one plain sentence with the number it wrote in it.
Why did the AI do that?
Every AI action expanded into the reason it happened: what it did, what it looked at (named, openable evidence), what rule applied (the clause, monospace), how confident (or an honest "not applicable"), and how to reverse it. Below: the check list (confirmed / not confirmed) and what the copilot proposes — numbered, editable, nothing actioned until a person approves.
Where does autonomy end?
The one dial an AI-ops product must put on screen: a stacked distribution bar of every decision by how well it was corroborated, with the evidence threshold drawn as a mark; the evidence ladder (countable tick meter, meaning, decisions, share, side); and three "If you move it" scenario cards that state the trade in the reader’s units (cleared share, queue delta, catches, reviews per catch) and MOVE the boundary when clicked.
What did the copilot find?
Recoverable-spend findings as a row of cards: a status badge naming the kind, a one-line title, the recoverable figure as the only coloured number (kpi-sm), a two-sentence summary and a primary + secondary action, under a header that states the programme and the total the cards add up to. The primary action is what a person signs; the secondary always opens the evidence.
Where was the context dropped?
One edge of a multi-agent run: sender → a metric pill on the arrow (context size, latency, tokens, fields present of required) → receiver; a context-comparison table (sent vs expected, status badge per field) that shows the required field that was not sent; the payload size against the workflow baseline as a Meter with a marker; and the edge’s 24-hour health beside it.
What changed while you were away?
The AI-ops briefing pattern: a numbered feed of things the copilot did or noticed — a one-sentence fact, a paragraph of why, the evidence it read, its confidence and two actions — with a conflict item that carries no confidence ("No verdict — a person decides"). Beside it, a "Needs a person" rail of what stopped and a "Derivation load" split of applied / sent to you / pinned by people.
Where does this number come from?
"Where the business stands" — four headline figures on ONE surface, each with an inline signed delta and the line that earns the number its place: source and freshness ("ERP · 2 min ago", "Derived from 61 open orders · 06:02"). Peers separated by hairlines, no value without provenance; the delta is secondary (arrow + sign, never colour alone).
Where did each field come from?
A record whose every field says where its value came from: derived (confidence meter + evidence kind), pinned (who and when — a pin out-ranks any derivation), conflict (two permitted sources disagree; nothing chosen) or empty ("No signal yet" — never a guess). Each state has its own glyph + text, and the footer counts the four so the reader knows how much of the record is machine-derived.
Why is the score 91?
A scored decision decomposed into its signals: one row per signal, a signed bar whose length is the contribution in points and whose direction is the sign — ink for positive, the warning rung for negative, each with a signed number so the sign never rides on colour. One shared zero-based scale; the footer states how many weaker signals were considered but not shown.
How close is each agent to its ceiling?
Every non-human spender on the card programme, sorted by share of ceiling: monospace agent id, autonomy band as a status badge (Auto / Review / Hold), a bar with the ceiling drawn as a vertical mark, the literal "$612 of $500", and Revoke. Over the ceiling paints the destructive rung and says so. Above it, three autonomy-band tiles total the envelopes per band.
Where did the run fail?
An execution waterfall for one multi-agent run: a stat strip where every number carries its comparison (duration vs p95, retries, tokens in/out, cost vs workflow average, failed at), the charts package’s Gantt at sub-second granularity with one row per agent span nested under its caller, and a context-flow chain that shows the hop where a required field was dropped.
Keep this, cancel that?
A keep-vs-cancel verdict side by side: two cards on the success/destructive washes with annual cost, a segmented seat-utilisation Meter and the deciding facts; a one-line reading of the comparison; “If you cancel” as four consequence stats; “How the copilot knows” as a numbered evidence list; and the refusal — the copilot prepares everything, a person signs.