Execution icons

The 37 execution icons in Agents, described

Action stepaction-step

An action step — one concrete thing an agent actually does in its loop.

Agent backagent-back

A chevron pointing back from an agent — rewind an agent to its previous step.

Agent queueagent-queue

An agent beside two queued items — the backlog of work waiting for an agent.

Agent sandboxagent-sandbox

An agent sandbox — an isolated place where an agent can run without touching anything real.

Agent stepagent-step

An agent with a chevron ahead of it — advance an agent by a single step.

Agent triggeragent-trigger

An agent beside a lightning bolt — the event that fires an agent into action.

Approval deniedapproval-deny

Approval denied — the human said no and the gate stays shut for the agent.

Approval grantedapproval-grant

Approval granted — the human said yes and the gate opens for the agent.

Blue team agentblue-team-agent

A blue team agent — paid to keep the system standing, a defending guard.

Code agentcode-agent

A code agent — writes the code, then reads it back to check it.

Deploy agentdeploy-agent

A deploy agent — ships the release and watches it land in production.

Escalate to humanescalate-human

Escalate to human — past what the agent should decide alone, handed up to a person.

Evaluate stepevaluate-step

Evaluate step — check whether the last step worked before taking the next.

Executorexecutor

The executor — the worker part of an agent that actually carries the planned steps out.

Fallbackfallback

A fallback — the backup path taken when the first way fails or times out.

Guarded actionguarded-action

A guarded action — the agent acts, but only inside the shield of its guardrails.

Interruptinterrupt

Interrupt — cut in right now, barge in and stop what the agent is doing.

Loop detectionloop-detect

Loop detection — the agent is going round in circles and has been caught.

Nudgenudge

A nudge — a tap rather than an order, a gentle poke to resume the work.

Red team agentred-team-agent

A red team agent — paid to break the system first, an attacking probe.

Sandbox agentsandbox-agent

An agent enclosed in a box — an agent running inside an isolated sandbox.

Sandbox escapesandbox-escape-alert

Sandbox escape alert — something is outside the box, and the box knows.

Scroll agentscroll-agent

A scroll agent — an agent that reads the way people read, scrolling down the page.

Shell agentshell-agent

A shell agent — an agent typing commands where the terminal prompt blinks.

Spend capspend-cap

A spend cap — the money stops here, a hard ceiling on the budget.

Step limitstep-limit

A step limit — only this many tries or iterations before an agent must stop.

Stuck agentstuck-agent

A stuck agent — still trying but not moving, stalled against a wall.

Summarise agentsummarize-agent

A summarise agent — long text in, a short digest out, the gist kept.

Task claimtask-claim

Task claim — this task is mine now, taken from the queue by one worker.

Task mergetask-merge

Task merge — two halves joined into one deliverable at the end.

Agent task queuetask-queue-agent

An agent task queue — what the agent will do next, in order of arrival.

Task splittask-split

Task split — too big for one, so now it is two subtasks handed out.

Task stealtask-steal

Task steal — idle hands take work from the busy pile to balance the load.

Task yieldtask-yield

Task yield — a task given back to the queue for someone else to pick up.

Test agenttest-agent

A test agent — runs the test suite so nobody has to remember to.

Triage agenttriage-agent

A triage agent — sorts the pile by priority before anyone opens it.

Turn limitturn-limit

A turn limit — the conversation has a budget of turns and then it stops.

In code, each is one import — import { ActionStep } from "@iconmind/react/icons/action-step" — and the same name in Vue, Svelte, Solid, Preact, React Native, Astro, Blade and Flutter.

Narrow it by tag