Agents · Batch

Not one answer.Every document, done.

Test an agent on one document. Then run the same agent across every document you have, with nothing rewritten.

Try it yourselfBook a demo
Free to start. No credit card.
→
AgentSNAPSHOT PINNED
→
Results → data source
DOCUMENTSUnderstood, scored and sourcedDATAStored for every way it's askedAGENTSScored, inside Hardrails
THIS PAGE DELIVERSConfidenceEvery result scored, at your real volume.

From assistant to automation

One answer helps a person.
Every document, done, changes the work.

An assistant answers one question at a time, so the work moves at the speed of the person asking. A batch runs the same agent across all of it.

AN ASSISTANTOne question at a time.
Is INV-20418 billed right?
No. Detention is $82 under the tariff, not $1,156.
And INV-20419?
That one is fine.
INV-20420?
0CHECKED2,400INVOICES THIS MONTH

A person asks, reads the answer and checks it. The backlog stays the same size.

AUTOMATION, ON MEIBELEvery invoice, every field.
run freight-audit-agent v7input march-invoices/ · 2,400 filessnapshot pinned · 50 at a timeoutput freight_audit_results
0CHECKED0FOR A PERSON

The same agent runs across the whole month. Every result is scored and sourced; people review only the exceptions.

ILLUSTRATION · SAMPLE PROJECT · PUBLIC DOCUMENTS · FREIGHT AUDIT

Set it up

Pick the agent, the files and the pace.
Then launch.

Launching a batch job in the Meibel console: a freight audit extractor agent selected, freight-audit-documents with 18 files as the input, a new output data source created from the job name, concurrency 20 and 2 retries per item.
  1. 1Pick a published agent.The version you tested is the version that runs.
  2. 2Point it at a data source.Each file runs as its own agent execution. Results land in a data source of their own.
  3. 3Set the pace.Up to 200 runs in parallel, retries per item, a timeout and a failure threshold.
OR IN PYTHONBatch jobs in the docs →
batch = client.batches.create(body=CreateBatchDefinitionRequest(
    name="eu-order-summaries",
    agent_id=agent.id,
    input_datasource_id="ds_abc123",
    user_message="Summarize the order history for this customer.",
))
# Run it on every item in the data source
execution = client.batches.execute(definition_id=batch.id)

Start with a folder you already have.

Refine the agent on one document. Run it on millions.

Try it yourselfBook a demo

Versioned runs

Change the agent mid-run.
The run doesn't change.

A batch pins the agent it started with: prompt, model, tools, schemas and scoring. Every result reflects the same setup.

THE AGENT
v6v7 · publishedv8 · draft edited
EDITED MID-RUN
MARCH BATCH
pinned to v7 · 2,400 of 2,400
APRIL BATCH
v8, once published
PROMPTv7 instructions
MODELYour choice, pinned
TOOLSSQL, search, read page
SCHEMA7 fields
SCORINGCorrect, faithful, grounded

Reviewed by exception

People see three invoices.
Not twenty-four hundred.

Every result is scored. You decide what a low score triggers, and the item waits with its source page beside the output, so the review takes seconds.

2,397 DONE · scored above your threshold, written to the data source
Delivery receipt · p.1 · fax
Arrived 08:10 · departed 1?:40
Billable hours2 or 3?
Governing rate$41.00 / hr
DecisionDispute

The departure hour is smudged. The agent asked rather than guess.

ConfirmFix the valueSend back
Rate sheet Q1 · p.3
Effective through March 31 · Amendment 2 from March 15
Billable hours2
Governing rate$41.00 or $44.00
DecisionDispute

The overlap needs a person to say which sheet governs this lane.

ConfirmFix the valueSend back
Invoice · p.1 header
Remit to: Carrier F Holdings LLC
CarrierCarrier F Holdings LLC
Contract partyCarrier F
DecisionHold

A possible new subsidiary. Confirm before paying.

ConfirmFix the valueSend back

SAMPLE RUN · SAMPLE PROJECT

Results become data

Every result lands as data.
The next agent starts from it.

Results land in a data source, tagged with the document they came from, ready for search, queries and the next agent.

The Meibel console: a completed batch run, 22 of 22 files done with none failed, each file listed with its status and the output it produced, with buttons to explore or export the results.123
  1. 1

    Every file accounted for.

    Completed, failed and total, so nothing slips through a run.

  2. 2

    Each file, its own result.

    Every output is tied to the file it came from, written to a data source.

  3. 3

    Explore it, export it, build on it.

    Search and query the results, or hand them to the next agent.

Real console · sample project · public bid documents

Next agent: build on the results →See every result in Data Explorer →Query it like any source →

Built on Meibel

Teams already ship on it.

We use Meibel to integrate generative AI into our SaaS product Specbooks+. We replaced thousands of lines of code and multiple other tools with better, measurable results.
Rob MurrayCEO, SpecBooks
Meibel is a foundational technology for our Generative AI practice. Our customers demand accountability and transparency from AI solutions, along with seamless integration into their existing operations and workflows. Meibel's platform delivers on all fronts.
Paul LiggettCEO, Cloud Eleven
Featured in

Read the case studies →

Questions

What runs in a batch?

One agent across many inputs: every invoice in a folder, every certificate from a supplier, every spec in a bid set.

Do I have to rewrite anything?

No. The agent you tested is the agent that runs.

What if I change the agent during a run?

The batch keeps the snapshot it started with, so every result reflects the same configuration.

What happens when an input fails?

It's retried automatically. Anything still failing can be rerun in a new run that targets only those inputs.

Where do the results go?

To an output data source, each result tagged with the document it came from, ready for search, queries and other agents.

Start with one workload. Scale efficiently.

Refine it on one document. Run it on millions.

Start free and test an agent on one real document. Then run one real process at your real volume and measure it against what your team already knows is right.

Try it yourselfBook a demo
PALETTE