← Go Pivot Edge · work
go pivot edge // agent runtimecapability demoSIMULATED

Watch an agent
actually work▮

Not a chat window with a new name. A machine that writes a plan, calls tools, hits a real error, recovers, and stops dead at anything it can't undo until a human says go.

or scroll — two more tasks in the runtime
01 · why this page exists

You've heard "AI agent" in nine decks this quarter. So has everyone we pitch. Almost none of them have watched one work — and the demos going around don't help, because they're screen recordings of the one run that behaved.

This scenario is invented. No client behind it, every company and number fictional, labelled SIMULATED wherever it appears. What it demonstrates is real: the runtime shape we deploy. A planner. Typed tools. A policy engine that refuses calls outright. Approval gates on anything external. And a meter, because tokens are money and anyone running these in production checks the bill.

What was broken in every agent demo we'd seen: the failure path was missing. Tools fail constantly — APIs 504, scrapers get blocked, search indexes time out at 3am. a demo that never shows an error is hiding the only part that matters.

So every run below contains one. Watch what happens after it. That's the product.

02 · the runtime — drive it yourself

Three tasks. Each one plans, works, breaks, recovers, and then stops at a gate and waits for you. Approve or reject — both endings are written, and they are not the same ending.

agentd 0.9.4sandboxSIMULATED
task
speed
tokens in0
tokens out0
est. spend$0.0000
tool calls0
wall clock00:00.0
agentd 0.9.4 · sandboxed replay · no live model calls
every company, number and tool result below is fictional.
select a task, pick a speed, hit RUN.
each run includes one genuine failure and one approval gate —
your approve/reject changes how it ends. both endings are real.
SIMULATED · scripted replay of runtime behaviour · fictional companies & figures · meter rates representative ($3.00/M in · $15.00/M out)
03 · what you just watched
plan
It writes intent before acting, and revises it mid-run where you can see it — the 2* line is the tell. A plan you can read is a plan you can audit.
tools
Typed calls with visible arguments and results. The agent can only do what a tool lets it do — that boundary, not the model, is the security model.
failure
The 504, the 403, the index timeout. Scripted here, because in production they're not scheduled. One retry, then a different route — never a third retry, never a crash, never quietly making the answer up.
the gate
External side-effects halt the run. And rejection is a first-class path: the agent files its work and stands down. It does not ask again. It does not sulk.
guardrails
The amber refusals in the side panel. Enforced by the runtime at the tool layer — not a paragraph in a prompt asking the model to please behave.
the meter
Tokens, spend, wall clock, call count. "What does a run cost" is the second question every operator asks. The first is "what happens when it breaks" — see above.
04 · where we won't put one yet

An honest list, because saying it out loud is the difference between having deployed these and having read about them.

irreversible
Anything that can't be undone doesn't run without a gate. Wire transfers, bulk deletes, contract sends. No gate, no agent — that one isn't negotiable.
wrong = expensive
Legal wording, medical anything, tax positions. An agent's failure mode is fluent and certain — precisely the wrong failure mode where a confident error costs real money or worse.
no test
Work you can't write an acceptance check for. If nobody can state what "correct" looks like, nothing in the loop can verify it. Fix that first; it's usually the real project.
latency-critical
A plan–tool–check loop takes seconds. Checkout doesn't have seconds. Those paths get deterministic code, not a model.
low volume
Twenty runs a month never pays back the harness. A checklist and a human is cheaper, and we will say so in the first call.
self-review
An agent as the only reviewer of its own output drifts, slowly and politely. Somewhere in the loop: a human, or a check that doesn't share the model's blind spots.

This list gets shorter every quarter. It is not empty yet, and anyone telling you it is empty is selling you the screen recording.

05 · your turn

The invoice chase above is invented — but it was assembled from the shape of things clients actually bring us. Describe your ugliest recurring workflow in three sentences. We'll tell you plainly whether it's an agent problem, a pipeline problem, or a checkbox in software you already pay for.

contact@gopivotedge.com →