It consults your systems.
MES, LIMS, CMMS, ERP, the warehouse. Each step reaches only the system it was given. Eleven of the fourteen steps call no model at all.
AGENTFLOW · AUTOMATE
Automate everything that can be automated.
Run, manage and improve the workflows that can't go wrong. The same path every time, checked at every step, with the cost, the speed and the proof on the record.
A REAL ONE, START TO FINISH
A batch fails at 2:40 a.m. Nobody is awake.
This is everything that happens next. AgentFlow reads from five systems, puts eleven agents to work at once, stops twice to ask a person, and writes back to four systems. Then it files the report.
14steps
8systems
3steps call a model
2decisions by people
4write-backs
38 minevent to report
RUN 7C41E2 · OUT-OF-SPEC-BATCH · V6completed 03:18
02:40:12
MES event. Batch B-24817 on line 3 fails viscosity at final QC.
- 02:40:1301Pull the batch recordMESruns as code
Recipe R-112, reactor 3B, 41 parameters, two operators.
- 02:40:1502Pull the lab resultsLIMScheck
Viscosity 4,310 cP against a limit of 4,200. Check passed: results present and signed by the lab.
- 02:40:1603Compare to spec and the last 30 runsWarehouseruns as code
Drifting upward for three batches. This one crossed the line.
- 02:40:1904Check the equipmentCMMScheck
Reactor 3B serviced May 30. Viscometer calibration valid to August, so the re-test branch is skipped. Check passed: instrument error ruled out.
- 02:40:2105Trace the raw material lotsERPruns as code
Seven lots went in. One is common to all three drifting batches: additive lot L-7731.
- 02:40:2406Find every batch that used itERP + MESfans out
Eleven batches, checked by eleven agents at once. Four in the warehouse, two scheduled, one shipped, four cleared by their own lab results.
- 02:40:4107Draft the root causemodelcalls a model
Lot L-7731 arrived with moisture at the top of its allowed range. Check passed: every claim cites a record.
- 02:40:5208Propose containmentplaybookcalls a model
Hold four batches. Quarantine what is left of the lot. Flag the shipped batch for customer notice. Plan frozen: 9f2c.
- 02:40:5309Ask the quality lead to approveSlacka person decides
No answer in 15 minutes, so it escalated. The plant quality manager approved at 03:02.
- 03:02:1010Place the holdsMES + WMSchanges something
Four batch holds and one lot quarantine written back. Exactly the frozen plan, nothing more.
- 03:02:1211Ask the plant managerSlacka person decides
"Two more batches are scheduled with this additive. Switch to lot L-7802, or pause line 3?" Answer at 03:14: switch.
- 03:14:3012Update the scheduleERPchanges something
Two production orders moved to lot L-7802.
- 03:14:3313Open the supplier corrective actionQMS + emailchanges something
Request SCAR-0419 opened with the evidence attached. The supplier's quality contact notified.
- 03:18:0214File the deviation reportQMS + Slackruns as code
Report DEV-2291 filed: findings, charts, every record ID. Posted to #quality-line3. Run written to the ledger.
runs as codecheckcalls a modela person decideschanges somethingfans out
Four minutes of work and thirty-four minutes of waiting for people. Done by hand, this is a morning's work for a quality engineer, and it starts when they arrive; by then the two scheduled batches have already run. Illustrative run drawn from the specialty-chemicals use case.
WHAT THAT RUN SHOWS
It did everything a careful person would, in that order.
Nothing in those fourteen steps is improvised. The path was fixed before the batch ever failed. Four things make that possible.
It works in parallel.
Eleven batches shared the suspect lot. Eleven small agents checked them at once, and their results folded into one table before the next step began.
It asks people, properly.
Once for a signature, once for a decision. When the quality lead did not answer in fifteen minutes, it escalated. The run waited, and would have waited for days.
It changes things, carefully.
Holds, a quarantine, a schedule change, a corrective action. Each plan was frozen before approval, so what ran is exactly what was signed.
FIVE PROMISES
What "can't go wrong" means, exactly.
Success for a workflow that matters is not one thing. It has to be right, the same every time, fast enough, worth what it costs, and provable afterward.
QUALITY
Correct
Every step is checked before the next one begins. It stops before it publishes a number it can't prove.
RELIABILITY
Repeatable
The path is fixed before the run starts. Same steps, same order, every time.
LATENCY
Fast
Independent work runs at once. Forty-two suppliers reviewed in 14.8 seconds on the benchmark workflow.
UNIT ECONOMICS
Worth it
You know what a run costs before you commit. As runs pile up, it proposes cheaper models, step by step.
AUDITABILITY
Provable
Every query, tool call, approval and cent goes on the record. Any number can be traced to its source.
THE WHOLE LIFE OF A WORKFLOW
Run it. Manage it. Improve it.
Writing a workflow is day one. AgentFlow is built for every day after that.
RUN
On a schedule, an event, or a sentence.
A trigger is one record: a schedule or a webhook, with input mapping and retries. One event can start ten workflows. Independent steps run at once, up to 256 at a time. State is checkpointed, so a run survives a restart and can wait days for a person.
MANAGE
Like anything else the business depends on.
Every change is saved as a new version. Every run shows what it cost, step by step. Approvals carry timeouts and escalation paths. The ledger keeps every run, so "what happened on March 3?" has an answer.
IMPROVE
It gets cheaper the more it runs.
After a few runs, AgentFlow studies them and proposes changes, usually a cheaper model on one particular step, with the expected effect on cost and speed at the same accuracy. You take them one at a time, or not at all.
HOW IT GETS WRITTEN
Written by the people who know the process.
The quality lead described that workflow in a paragraph. AgentFlow wrote the steps, picked the tools, and showed her the draft. She changed it the same way, a sentence at a time, with no code, no engineering ticket, and a new version saved every time.
PROJECT · PLANT QUALITYplain English in, a versioned workflow out
YOU
When a batch fails final QC, investigate it, contain it, and open a corrective action with the supplier. Ask the quality lead before anything is held.
AGENTFLOW
Drafted 14 steps across MES, LIMS, CMMS, ERP, WMS and QMS. Three call a model, two wait for a person, four change something. Review them?
FIVE VERSIONS LATER, IN THE SAME CHAT
YOU
If the viscometer's calibration has expired, re-test before doing anything else.
A rule stated in one sentence becomes a branch the run has to take.
USE CASES
Three processes, automated.
Each one told step by step: what starts it, what runs, what is checked, and what arrives.
The out-of-spec batch
For quality. The run on this page, with who owns it and what it replaced.
Read moreThe close, explained
For finance. Every account over the threshold, explained from the journals, with the document IDs.
Read moreSupplier risk, every Monday
For procurement. Forty-two suppliers reviewed at once, one ranked summary in Slack.
Read moreQuestions about Automate
What happens when a step fails?
The run stops at that step and says which one and why. It does not carry on with half an answer. Steps carry checks: the shape of the output, a file that must exist, a total that must reconcile. They also get a retry, prompted with exactly what was missed.
Can a person approve something before it happens?
Yes. A workflow can pause for a named person and reach them in Slack, by email or in the product. Approve and it continues. Decline and it stops. Approvals carry timeouts and escalation paths, so a run never hangs on someone who is asleep.
Can it change things in our operational systems?
Yes, carefully. A step that changes something plans the change, freezes that plan with a fingerprint of its inputs, routes it for approval, runs only on approval, and then writes an append-only record. Operational sources are read-only unless you say otherwise.
Can we predict what a run will cost before we commit to it?
Yes, and that follows from the path being fixed. The steps are known in advance and most of them don't call a model, so cost per run is a number you can see while you write the workflow and hold to afterward.
Do we need engineers to maintain the workflows?
No. The people who own the process change it in chat, in a sentence, and every change is saved as a new version. AI engineering keeps control of the models, the tools and the boundary, and reviews whatever it chooses to review.
TALK TO US
Bring us the process you worry about most.
We will show you how it would run on AgentFlow: which systems it would touch, where it would stop to ask, and what each run would cost.
Read the docsThanks.
An actual person replies, usually within one working day.