AI agents you can hold to account
We design, build and run agents that take on real work: customer enquiries, phone calls, back-office operations. Every one ships with guardrails, an evaluation suite and monitoring, so you can see how it behaves rather than hope.
Clear prices in writing · senior engineers only · NDA as standardWhere agents earn their keep
Not moonshots. Bounded, repetitive, high-volume work where the rules can be written down, and where a missed step costs real money or a real customer.
Enquiry triage
Reads every enquiry as it arrives, looks the sender up, answers what it is authorised to answer and routes the rest with a drafted reply attached. Nothing waits until morning.
Call answering
After-hours and overflow calls answered by voice: callers qualified, appointments booked into your calendar and CRM, emergencies escalated by rule. No voicemail black hole.
Back-office processing
Invoices, orders and documents moved between systems with validation at every step. Discrepancies are flagged to a person, never silently passed through.
Research and reporting
Internal agents that monitor sources, summarise changes and draft the weekly report for a human to sign. The drudgery goes; the judgement stays yours.
Patterns, not promises: every build starts from your actual workflow in discovery.
Anyone can demo an agent. We prove ours behave.
A demo shows the happy path once. Production means the ten-thousandth interaction, the malicious one and the weird one. Three disciplines make that safe to rely on.
Guardrails
Every agent has written limits of authority: what it may say, spend, access and decide. Anything outside those limits stops and escalates to a named person. The limits are agreed with you in discovery and enforced in code, not in the prompt.
Evaluation
A test suite of real scenarios, including the hostile and the absurd, runs before every release. Scores must clear agreed thresholds or the release does not ship. You see the results, not a reassurance.
Monitoring
Every action is logged with a full trace, so any decision can be reconstructed afterwards. Escalations and failures are reviewed on a schedule, and you get the same view we do.
Built to the standard we audit to. Before any agent goes live it passes the same senior review we sell as the Vibe Code Audit: security, failure handling, data posture. We would not sign off anyone else's shortcuts, so we do not ship our own.
About the auditThe same discipline, on the phone
A voice agent speaks for your business in real time, which is exactly why it gets the strictest guardrails we write.
Good evening, you have reached the out-of-hours line. I am an automated assistant and this call is recorded. How can I help?
My boiler has stopped working and we have no heating.
I can book an engineer for tomorrow at 8:30. Could I take your postcode?
Thanks. Also, I think I can smell gas near the meter.
That is an emergency. Please call the National Gas Emergency Service on 0800 111 999 now. I am alerting the on-call engineer as well.
What a Zegaware voice agent does
Already chosen your agent? We install it properly.
If an off-the-shelf open-source agent fits your needs, you do not need a bespoke build. We install, harden and maintain the two leading ones, on your server or ours.
OpenClaw
The viral open-source agent that lives in your messaging apps. We install it isolated and hardened, connect your channels and keep it updated.
OpenClaw installation →Hermes Agent
The self-improving agent from Nous Research, with persistent memory and skills that compound. We install, maintain and migrate from OpenClaw.
Hermes Agent installation →Four steps, no leap of faith
Agent builds vary too much for a fixed timetable, so we do not promise one. The shape is always the same, and you can stop after any step.
Map the workflow
We sit with the people who do the work now. Authority limits, escalation rules and success measures are written down and agreed. Fixed price quoted.
Run it alongside
A working agent handles real cases in shadow mode while your team keeps control. You judge it on its trace log, not on a demo.
Prove it behaves
Guardrails enforced in code, the evaluation suite built out from pilot cases, failure handling tested. Senior review before anything goes live.
Operate and tune
Live with monitoring, scheduled reviews and re-evaluation on every change. You see the same dashboard we do.
Priced by the workflow, not by the hour
Agent builds vary too much for a price list, so we will not pretend otherwise. The shape of every engagement is the same; the figures are confirmed in writing after discovery, before any build starts.
A short, scoped piece of work with a written output: the workflow map, authority limits and a firm quote for the build. If we think an agent is the wrong tool, this is where we say so.
Quoted after discovery: a fixed price when the scope is clear, or logged time billed monthly when the work will evolve, with the pilot and hardening steps included. We recommend the model that fits and confirm scope, price and any deposit in the proposal.
Hosting, monitoring, scheduled reviews and re-evaluation on every change. Plain-English terms; you are never locked in to keep the agent you paid for.
Asked before every build
What happens when the agent gets it wrong?
It fails loudly, by design. Anything the agent is unsure about, or that falls outside its written authority, stops and escalates to a named person with the context attached. Every action is logged, so when something does go wrong you can reconstruct exactly what happened and we can fix the cause, not the symptom. What we do not build is an agent that guesses quietly.
Will it say something embarrassing to a customer?
This is the main risk of customer-facing agents and we treat it as an engineering problem, not a hope. The agent works from your approved knowledge and within written limits on what it may say and offer. The evaluation suite includes hostile and absurd inputs, and releases do not ship until they pass. No system makes the risk zero; ours makes it measured, bounded and visible.
What does it integrate with?
Whatever the workflow touches: phone systems, shared inboxes, CRMs, calendars, accounting packages and internal databases. Integration work is part of the build, not an extra. If a system has no sensible interface, we say so in discovery rather than discover it for you mid-build.
Do we need our data and processes sorted out first?
No. Discovery works with the process you actually have, including the undocumented parts that live in one person's head. Messy reality narrows what the agent should be trusted with at first, and that is reflected in its authority limits rather than ignored.
Where does our data go, and what about GDPR?
Data flows are mapped in discovery and agreed before the build: which systems the agent reads, what it stores, for how long, and which model providers see what. We work under NDA and a data processing agreement, log access, and configure retention to your policy. Callers and customers are told when they are talking to an agent.
Someone else built us an agent. Can you take it over?
Yes, and we start the same way we start with AI-written code: an audit. We review what exists against the guardrails, evaluation and monitoring standard described above, give you a scored picture of where it stands, and quote for bringing it up to that standard or running it as is.
Do you work fixed-price or by time?
Both, and we recommend the one that fits the work in writing. A fixed price suits a clearly scoped build: you get certainty and we carry the estimate. When the work will evolve, or you would rather just send tasks as they come, we work to logged time billed monthly, so you pay for the time the work actually takes, with no quote to wait for on every change and no penalty when scope grows. Agent work is new enough that forcing a fixed price on exploratory work means either padding it or absorbing overruns, neither of which is fair to you, so time-based is often the honest choice there. Whichever model we use, the rate and the shape are agreed in writing first.
Tell us about the workflow
We reply within one working day with how we would approach it and a fixed fee for discovery. No call required unless you want one.