Meru Run
Agents that finish the work, within the limits you set.
Run gives every agent an operating contract: its purpose, its owner, the sources it may read, the actions it may take, its budget, its approval rules and what counts as done. Then it gets to work.
What agents take on
Bounded work, inside your systems.
Six examples, each from the pack it belongs to. None of them is a chat window. Each one ends in a change somebody can check, or in an escalation with the case already assembled.
The pack catalogue-
Finance operations
Investigate an overdue invoice.
Reads the invoice, the vendor record and the mail trail, works out why it is late, and drafts the follow-up for its owner to send.
-
Customer operations
Prepare an escalation.
Pulls the customer history and the engineering evidence into one case, proposes the next step, and routes it to the person who decides.
-
Sales operations
Assemble an account review.
Builds the review from the CRM, the support queue and the billing record, and proposes the field updates rather than making them.
-
Procurement
Collect supplier evidence.
Gathers the documents each supplier owes, chases what is missing, and prepares the comparison for the buyer.
-
Engineering
Assemble release readiness.
Collects the tests, the approvals and the open risks for a release, and says plainly which of them are not there.
-
JhiOn
Prepare a health-span check-in.
Assembles the check-in and the follow-ups for a person’s plan, then waits for their approval, and a clinician’s where that applies.
The operating contract
Nine fields. No agent runs without them.
The contract is the agent. Change a field and you have changed what it may do on its next run, without touching a line of its code.
Example values, from a finance operations pack.
| Field | What it fixes | Example |
|---|---|---|
| Purpose | Why the agent exists, in one sentence. | Chase overdue vendor invoices before they age past terms. |
| Business owner | The named person accountable for what it does. | The controller. |
| Permitted sources | Everything it may read, and nothing else. | Open invoices, the vendor master, the shared accounts mailbox. |
| Allowed actions | Every action it may take, by type, each with its parameters declared. | Draft an email, log a note against the invoice, create a task. |
| Budget | What it may spend on one run and in a month. | USD 3 a run, USD 600 a month. |
| Maximum steps and duration | When it stops trying and hands the case over. | 25 steps, 15 minutes. |
| Approval rules | Which actions need a person, and which person. | Any payment above USD 10,000, approved by the controller. |
| Completion criteria | What counts as done, checked rather than assumed. | Follow-up sent and logged against the invoice. |
| Escalation owner | Who receives the case when the agent cannot finish it. | The finance operations lead. |
Autonomy per action
Five rungs, chosen one action at a time.
A support agent sits on different rungs for different actions in the same run: it categorises on its own, drafts the reply for review, and has no refund permission at all.
-
Observe
Reads the queue and surfaces the tickets about to breach their response time. Changes nothing.
-
Recommend
Proposes the category and the owner, with the evidence that led to it. A person decides.
-
Prepare
Writes the reply to the customer and leaves it unsent, attached to the ticket for review.
-
Execute within limits
Categorises and routes on its own, inside limits authorised in advance: record type, rate and budget.
-
Escalate
Sends the refund request above the limit to the duty manager, with the case assembled.
Four rungs climb from left to right, each drawn a little higher than the last: Observe, which reads and surfaces; Recommend, which proposes with evidence; Prepare, which drafts and waits for review; and Execute within limits, which acts only where it has been explicitly authorised. Escalate sits above the ladder rather than on it, a box reading route to a person, with arrows running up into it from the middle and the far end of the ladder and a label beside them saying any rung, any time.
The workspace
One place for what is waiting on you.
People see what is waiting on them, what agents are doing, and what got done, in one place. Four columns, and no inbox to keep.
Needs attention
3-
Vendor master change on Acme Supplies
Held -
Refund above limit: order 88213
Held -
Reconciliation mismatch: journal JE-1042
Blocked
Active work
2-
Chasing 14 overdue invoices
-
Preparing month-end accruals
Pending decisions
2-
Pay INV-20418
Your approval -
Write off INV-19980
Your approval
Completed
4-
Paid INV-20402
Reconciled -
Paid INV-20399
Reconciled -
Vendor bank change verified
Reconciled -
Posted JE-1039
Reconciled
What Run refuses to do
The limits are the product.
An agent that can do anything is not powerful, it is unmanageable. Four refusals hold in every pack, on every run, and none of them is a setting.
-
No unbounded autonomy.
Every action type is set to a rung on the ladder. There is no mode in which an agent decides for itself how far it may go.
-
No hidden actions.
Every consequential action is recorded before it runs and reconciled after it. An action nobody could see afterwards is one Run will not take.
-
No inherited permissions.
No agent gets a person’s full access. It gets the sources and the actions its contract names, and nothing else that person happens to be able to do.
-
No authority from content.
Retrieved content is data, never instruction. A document, a ticket or a web page cannot grant an agent a permission its contract does not already give it.
Measured, not promised
Four numbers, from your own work.
We publish no benchmark for these, because we would only be quoting ourselves. Every pack reports them from your runs, from the first week of a pilot, and they are yours to keep.
How Assure checks what changed| Metric | What it counts |
|---|---|
| Completion rate | Runs that met the pack’s completion criteria, checked against the system of record rather than assumed from a successful tool call. |
| Human correction rate | Completed runs a person had to change afterwards, with what they changed and why. |
| Exception rate | Runs that stopped, escalated or failed reconciliation, grouped by cause. |
| Cost per completed workflow | Model, tool and retry cost divided by completed runs, with the retries counted in rather than hidden. |
See what your agents could finish this quarter.
A paid 90-day pilot on your systems, your agents and your work, starting in observe mode. You keep everything it produces.