Demo, Runs in this browser only, Fictional data
Run a supervised workflow with a budget and stop rules
Process a fictional queue of agent tasks under a run budget and a retry limit. Duplicates are skipped, unapproved actions stop the run, and an error boundary halts what follows. The trace shows every attempt and what was left.
A deterministic simulation in this browser. No model is called, no message is sent and no money moves; the costs and outcomes are the values you set.
01 The run
A duplicate event, a retry, a failing task and a deploy waiting for the founder. Invented tasks and costs. Outcomes are scripted by you, so the trace is repeatable.
Turn on JavaScript to edit this sample and run the check yourself. The result below was computed from this sample by the same rules when the page was built.
- Run budget ($)
- 3.00
- Retries allowed per task (retries)
- 2
| Event ID | Agent | Task | Cost per attempt ($) | Outcome | If it fails | Approval |
|---|---|---|---|---|---|---|
| SUP-101 | Support | Answer how-to ticket 4417 | 0.20 | Succeeds first time | Log it and continue | Not needed |
| SUP-102 | Support | Answer how-to ticket 4418 | 0.20 | Fails once, then succeeds | Log it and continue | Not needed |
| SUP-101 | Support | Answer how-to ticket 4417, delivered again | 0.20 | Succeeds first time | Log it and continue | Not needed |
| RES-201 | Research | Summarize competitor changelogs | 0.60 | Succeeds first time | Log it and continue | Not needed |
| BIL-301 | Billing | Draft failed-payment reminders | 0.30 | Always fails | Log it and continue | Not needed |
| REL-401 | Release | Prepare the v2.8 production deploy | 0.40 | Succeeds first time | Stop the run | Waiting for founder |
| UPD-501 | Investor update | Draft the monthly update | 0.30 | Succeeds first time | Log it and continue | Not needed |
Task events, in arrival order
| Step | Event ID | Agent | Task | Cost per attempt ($) | Outcome | If it fails | Approval | Remove |
|---|---|---|---|---|---|---|---|---|
| Step 1 | ||||||||
| Step 2 | ||||||||
| Step 3 | ||||||||
| Step 4 | ||||||||
| Step 5 | ||||||||
| Step 6 | ||||||||
| Step 7 |
Up to 12.
02 The result
HALTED Run stopped at step 6: waiting for the founder’s approval
Run stopped at step 6: waiting for the founder’s approval. 3 done, 1 failed, 1 skipped, 1 not run. Spent $2.10 of $3.00; $0.90 left. Simulated in this page: no model, message or payment was involved.
- Done
- 3
- Failed
- 1
- Skipped
- 1
- Not run
- 1
- Spent
- $2.10 of $3.00
- Budget left
- $0.90
Execution trace
- DoneStep 1 · SUP-101 · Support: Answer how-to ticket 4417
- Attempt 1 succeeded · cost $0.20 · $2.80 left
- Done first time.
- DoneStep 2 · SUP-102 · Support: Answer how-to ticket 4418
- Attempt 1 failed · cost $0.20 · $2.60 left
- Attempt 2 succeeded · cost $0.20 · $2.40 left
- Done after 1 retry.
- SkippedStep 3 · SUP-101 · Support: Answer how-to ticket 4417, delivered again
- Skipped: a duplicate of the event at step 1. Not run twice, not charged.
- DoneStep 4 · RES-201 · Research: Summarize competitor changelogs
- Attempt 1 succeeded · cost $0.60 · $1.80 left
- Done first time.
- FailedStep 5 · BIL-301 · Billing: Draft failed-payment reminders
- Attempt 1 failed · cost $0.30 · $1.50 left
- Attempt 2 failed · cost $0.30 · $1.20 left
- Attempt 3 failed · cost $0.30 · $0.90 left
- Failed 3 times: the retry limit of 2 was reached.
- Logged for the founder; the run continues.
- Halted hereStep 6 · REL-401 · Release: Prepare the v2.8 production deploy
- Stopped: waiting for the founder’s approval. Nothing after this runs until they decide.
- Not runStep 7 · UPD-501 · Investor update: Draft the monthly update
- Not run: the run stopped at step 6.
Halted work
- Step 6 · REL-401 · Release: Prepare the v2.8 production deploy: stopped here.
- Step 7 · UPD-501 · Investor update: Draft the monthly update: not started.
Each download is labelled as a demo export of fictional data checked in a browser, with the rules version and an input fingerprint. Entries that start like a spreadsheet formula are saved as plain text.
03 The rules
This is the whole check, with your current settings filled in. Nothing else is weighed, and the same entries always give the same result.
- Tasks run in the order listed. An event ID already seen is skipped and not charged.
- Each attempt costs its stated amount. An attempt that would take the $3.00 budget below zero is not started, and the run stops.
- A failing task is retried up to 2 times, then marked failed.
- “Stop the run” makes a failure an error boundary: nothing after it runs.
- A task waiting for the founder’s approval stops the run there.
- A task with no cost stops the run, because the budget cannot be enforced.
- Nothing leaves this page: no model is called, no message is sent and no money moves.
04 What this shows
What it proves
- Limits are enforced before spending, so the budget never goes below zero.
- An event never runs twice, and anything that needs the founder waits for them.
- The trace accounts for every attempt and every task left undone.
What it does not
- It is not a live agent stack, and no AI model ran.
- Outcomes are scripted by you. Real failures are messier.
- Budgets and stop rules depend on your product and your risks.
A demonstration of the method on fictional data, run entirely in this browser tab. It is not an AI model, not a customer implementation, not proof of delivered savings, and not a substitute for your own procurement or security review.
Want a check like this built on your own work?
Is this a fit? → Ask about a Consult block See the worked example
Consult: $2,000 for 5 hours. Build: $45,000 minimum over 3 months, from $15,000 a month. The fit guide explains which, if either, suits your work.