Agentic Systems

A team of agents, built around your business.

One agent does one job. In a system, agents brief each other. One asks another to check something, gets the answer and carries on, so what reaches you is finished and already checked. We shape it around how your business already runs, and we run our own consultancy on one.

Team chat · #agents · shared thread

JD
Inbox
Accounts
Reporting
You

Inbox agent

@Accounts Harbour Studio replied to our chase. They say INV-1042 was paid on Tuesday. Can you check?
Accounts agent is working

Illustrative · agents briefing each other, one decision for you

What’s in the system

More than one agent. A team with a way of working.

An agent does a job. A system decides which agent does what and who signs it off. These are the parts we put together, in whatever shape suits the business.

Intake

Reads everything that comes in

Inboxes, forms, meeting recordings and the systems you already run. Each item is matched to the right customer or job and turned into work, with a draft attached where one makes sense.

emailRe: revised scopedraft ready
meetingTuesday check-in3 actions
formWebsite enquiryreply drafted
accountsINV-1051 queriedheld back

Specialists

An agent per job

Accounts, sales, reporting and whatever else your week needs. Each owns its tools and its rules. Add one without rebuilding the rest.

Teamwork

Agents brief each other

One agent asks another to check a payment or update a report, then carries on with the answer. Every exchange is in a thread you can read.

Approvals

In the tools you already use

Prepared work shows up as a card in Teams, Slack or email. Approve it there and the job closes.

The run log

Every run on the record

Inputs, outputs and cost are logged on every run. When something looks off, you can replay it.

Learning

Edits become rules

Your changes to drafts are kept. Each week they’re turned into rules you accept or bin. Accepted rules apply from the next run.

Memory

A living file on every customer

Every email, meeting and job is kept against the customer, so any agent can pick up where the last one left off.

A typical day

You see the decisions. The groundwork is done.

  1. 06:00

    Overnight intake

    Emails, forms and messages from overnight are sorted and matched to the right customer.

  2. 06:30

    Drafts ready

    Replies, chases and follow-ups are drafted before anyone logs in.

  3. 07:30

    Morning brief

    One message: what needs you today and what the agents handled.

  4. All day

    Approvals

    Prepared work shows up in Teams, Slack or email. Approve it where you are.

  5. Every 10 minutes

    Watching

    Systems are checked, and sent mail is matched to its job so nothing gets chased twice.

  6. Monday 07:30

    The weekly report

    The numbers pulled together, with the commentary written.

  7. Friday

    Learning

    A week of your edits becomes proposed rules for you to accept.

Examples

Every system looks different.

Each one is built from jobs like these, in whichever tools the team already lives in.

All examples

How we build one

One job first. Then the rest, on top.

01

Map the week

We sit with the inboxes, the calendar and the spreadsheets and list the work that recurs. Then we pick the loop worth automating first, usually the one that eats the most hours.

02

First loop live

An intake agent, one specialist and the approval step, running on your data in the tools you already use. Real work, signed off by a person, from the first week.

03

Add specialists

Each new agent works from the same record, so it plugs into what’s there. Finance after sales, reporting after finance, at whatever pace suits the team.

04

Run and tune

We watch the run log, accept or reject the rules the system proposes, and move agents to better models as they ship. You stay on the decisions.

The rules we build by

Autonomy where it's safe. A person where it counts.

Nothing leaves without a person

Sending, paying and deleting sit behind an approval. We build that block into the code, so no prompt can talk its way past it.

Every run is on the record

Inputs, outputs and cost are logged on every run. When something looks off, you can replay it.

You own the data

It stays in systems you control: your Microsoft 365, your Google Workspace or a database we set up in your name. The AI models underneath stay swappable.

Your edits become rules

Corrections are kept and reviewed weekly. The good ones become standing rules the agents follow from the next run.

run log · today

example
06:00:02inbox41 emails sorted · 9 replies drafted
06:00:40accounts3 overdue invoices · 2 chases drafted, 1 held
06:10:05monitoringpayments failing · cause found · owner told
07:30:00reportingweekly numbers built · commentary written
08:41:12salesnew enquiry researched · reply drafted
09:02:51accounts2 chases approved by Jo · sent
09:14:03inbox“Re: revised scope” approved · sent
11:20:44meetings3 actions from check-in · follow-up drafted
14:05:19support6 tickets answered · 1 passed to a person
16:40:10salesreply approved · lead moved to call offered
Fri 17:00learning1 rule proposed from 4 edits
06:00:02inbox41 emails sorted · 9 replies drafted
06:00:40accounts3 overdue invoices · 2 chases drafted, 1 held
06:10:05monitoringpayments failing · cause found · owner told
07:30:00reportingweekly numbers built · commentary written
08:41:12salesnew enquiry researched · reply drafted
09:02:51accounts2 chases approved by Jo · sent
09:14:03inbox“Re: revised scope” approved · sent
11:20:44meetings3 actions from check-in · follow-up drafted
14:05:19support6 tickets answered · 1 passed to a person
16:40:10salesreply approved · lead moved to call offered
Fri 17:00learning1 rule proposed from 4 edits

Illustrative · one day of a typical build

Questions

What people ask before they start.

  • They draft. They don’t send. Sending needs a person to tap approve, and on the systems we build that block is in the code, so a clever prompt can’t get round it. Every run is logged with what went in and what came out.

Free 30-minute call

See one running.

On a free call Neil can show you the system NLP Consultancy runs on, then map which job in your business we’d hand to an agent first.