Managed agents · Built & governed
Managed agents
Software that does the work, with credentials, limits and someone accountable when it gets it wrong.
UK-based · Open source stack · You keep the keys
An agent holds credentials to your systems and can act without a person pressing go. That is the reason this is a service rather than a download. Before anything is connected we agree permitted data, minimum access, which actions need human approval, what happens when it fails, and which model providers are allowed.
Definition
What an agent actually is
A chatbot answers. A script follows steps. An agent is given an outcome, decides which tools to use, does several steps in a loop, and reports back.
A goal, not a script
You give it an outcome in plain language. It works out the steps, rather than following a sequence someone hard-coded.
Tools and credentials
It connects to your inbox, spreadsheets, CRM, website or internal systems, and it can read, and sometimes write, on your behalf.
Permission to act
It runs without anyone pressing go. That is the point, and it is exactly why it needs limits, logs and a named owner.
That combination, credentials plus autonomy, is the whole reason the job exists. The moment software can send, delete or pay, somebody has to own what it does.
Three products
Audit, pilot, then managed
Buy the audit if you already have something running. Buy the pilot if you have a job that eats time. The managed service is what keeps it working afterwards.
Agent audit
For teams already using AI and unsure what it can reach.
one week, no obligation
- Inventory of every agent and automation, and who owns it
- Credentials, scopes and what it can read, write or delete
- What data leaves the building, to which provider
- What it can do unsupervised, and what breaks when it is wrong
- Model spend, retry loops and hidden running costs
- Two-page report, risk register and a go/no-go walkthrough
Pilot
One workflow, one channel, thirty days of real use, measured.
then from £150/mo managed
- One specific job, agreed before we build
- One channel: Telegram, Slack, WhatsApp or email
- Two integrations, scoped tightly
- Read-only first, with approval for anything consequential
- Success measures agreed up front, and reported at the end
- If it misses the measures, you stop
Managed
The ongoing product. The agent keeps working, or we are the ones who fix it.
scoped to what you run
- Monitoring and alerting on failures, not just uptime
- Prompt, model and integration updates
- Cost control and usage reviews
- Incident response when it misbehaves
- Quarterly review with changes agreed in writing
- You keep the configuration, so you can leave with it
Model usage costs are separate and agreed in advance. The audit is fixed price. Pilots are fixed scope. Managed fees reflect what you actually run, not a per-seat count.
Guardrails
Access you can justify
Most AI builders can add a tool. Fewer can tell you exactly what it can reach, what it did, and how to stop it.
Least access, on purpose
An agent starts read-only. Write access is added one system at a time, with the smallest scope that does the job.
Every action logged
What it read, what it changed, what it sent, and who approved it. If nobody can see it, it does not ship.
A kill switch that works
A documented way to stop it in under a minute, tested before go-live, because agents fail at the worst time.
Example workflows
Work an agent can take off a small team
These are examples. Yours gets scoped to your systems, your permissions and the one job that wastes the most time.
Support triage
Reads the inbox, drafts replies, routes the awkward ones to a human, and notices when a customer is angry.
Order and admin handling
Takes changes from email or a form, updates the order book, drafts the supplier order and flags anything unusual.
Scheduled reporting
Pulls the numbers together on a schedule and sends a digest, with the caveats a human would have added.
Document Q&A
Answers staff or customer questions from your own policies and documents, and says when it does not know.
Lead qualification
Asks the right questions, scores the enquiry and books the good ones, leaving the judgement calls to you.
Monitoring with judgement
Watches your services, correlates the noise and pings the team with context rather than a wall of alerts.
What we build on
Open source, and portable
No per-seat licence, no platform you cannot leave. The configuration is exportable and the stack is one you could hand to another engineer.
- OpenClaw agent platform
- Open-weight models where they fit
- Containers on Linux
- Postgres for state
- Your infrastructure, or dedicated managed
Where it fits
Teams with more work than hours
Best where there is repeatable work that eats time and where a mistake has a real cost: triage, admin, reporting, the same questions every week.
- Agencies
- Co-ops and social enterprises
- Publishers
- Professional services
- Ops-heavy small teams
Talk it through
Start with the audit, or with a pilot.
Tell us what you already run, or which job you want handled. We will say which of the three fits, and quote it before anything is connected.
Send an enquiry at any time. We will respond as soon as we can and agree a practical next step with you.