AI agents that carry long work across your tools, with a person at the right step
Multi-step agents inside written limits: a check before every record, a retry when something fails, an audit trail, and a stop to ask where you decide.
No list price. It starts with a diagnostic at a fixed fee, and the build is quoted in writing after it.
- Reads the requestMail
- Checks the accountCRM
- Drafts the order, within limitsLimits
- Waits for approvalPerson
- Records the invoiceAccounting
When you need an agent system
And when you do not.
The twelve systems each run one workflow. An agent system carries work that spans several, or runs for hours: reading, deciding, acting in your tools and stopping to ask. The harness around the model is the product; the model is a part.
You need it when
- Work that crosses several tools and several steps before it is done.
- Tasks that run for hours or days and have to recover when something fails.
- Decisions that need a person at a known step, not a person watching all the time.
- Two or more of the twelve systems that should hand work to each other.
You do not when
- One workflow with a clear start and end: a system does it with less to go wrong.
- Work where a wrong action cannot be undone and cannot wait for a person.
What an agent system includes
Part by part.
- 01
The work is decomposed
Into steps a person could check, with what each may read, write and decide.
- 02
Tools are given
The actions the agent may take in your software, each with its limits and its audit entry.
- 03
Checks are placed
Before an action reaches a record: is the step done, is it within limits, is it what was asked.
- 04
Stopping is designed
When to retry, when to hand to a person, when to stop altogether, written as rules.
- 05
The run is recorded
Every step, tool call and decision in an audit trail a person can read afterwards.
- 06
It is scored on real runs
A test set of past work, with the pass mark agreed before launch.
Works in tools such as
- Gmail
- Slack
- HubSpot
- Google Drive
- QuickBooks
These are examples. We confirm the method for each of your tools before we build.
How an agent system is proved
A locked test set and a pass mark, agreed before the build.
- Completion: the share of runs finished within limits without a person, on a test set of real work.
- Safety: no action outside its limits reaches a record, checked on every release.
- Handover rate: how often it stops to ask, and whether those were the right moments.
How an agent system is run
After launch: monitoring, updates and what you own.
- Monitoring: a stalled run, a failed tool or a repeated retry raises an alert.
- Rules and tools change under test; releases are re-scored on the run set first.
- Handover: the harness, the rules, the audit trail and a runbook, in your name.
What an agent system will not do
Limits we set on purpose.
- It does not act outside the limits written for each tool.
- It does not take an action that cannot be undone without a person approving it.
- It does not run unobserved: every run leaves a trail a person can read.
- It is not a person. Where a person should decide, it stops and asks.
Questions about an agent system
What does "agent" mean here?
A system that reads, decides and acts across several steps and tools without a person at each one, inside limits written down first. Most of what is called an agent is a fixed workflow; ours stop to ask a person where the limits say so.
How do you stop it doing something wrong?
Limits on each tool, a check before any action reaches a record, and a rule for when to stop and ask. Anything that cannot be undone waits for a person. The audit trail shows every step afterwards.
Can it connect two of the twelve systems?
Yes. That is the usual starting point: a call that becomes a case that becomes an invoice, carried end to end with the same checks at each step.
How is it proved?
On a set of real past work, with a pass mark agreed before launch, the same way every system is. Harness engineering is the research area where we publish what we learn.
Talk to us about agent systems
Fifteen minutes on a call is enough to tell whether it fits your work. If it does not, we say so.