AI & agents

We design, deploy and operate AI agents that do real work inside a business: reading, deciding, acting across your systems, with the engineering discipline that production software demands.

Agents in production.

Not demos. Agents with access controls, audit trails, evaluation suites and a human in the loop wherever a decision warrants one.

AgentsAutomationEvaluationObservability

What we deploy

Operations agents

Invoices, claims, orders, scheduling and reconciliations: the high-volume, rules-heavy work that ties up skilled people.

Service agents

Customer and member enquiries across email, chat and voice, resolved end to end or handed to a person with full context.

Monitoring & maintenance agents

Watching logs, alerts, queues and data feeds around the clock, investigating anomalies and paging a person with the evidence attached. Keeping software healthy too: bugs reproduced and fixed, failing builds repaired, and dependency and security updates raised as tested changes for an engineer to approve. Support

Research & reporting agents

Monitoring sources, assembling evidence and drafting reports and briefs on a schedule.

Sales & outreach agents

Prospect research, qualification and personalised outreach, within the rules you set.

Internal copilots

Assistants grounded in your own documents, systems and data, with permissions that match your organisation.

Visual content pipelines

Art-directed image and video libraries generated, checked and versioned at catalogue scale. More

Agent-ready products

MCP servers that let customers' assistants and agents use your software. More

An agent at work

Accounts payable agentRun 2,418 · illustrative
14Invoices in
12Bills drafted
2Exceptions resolved
1Handed to a person
$0.58Model cost
  1. Intake

    Supplier invoices collected from the accounts inbox.

    inbox.searchfiles.fetch ×14
    14 received
  2. Read

    Supplier, line items, totals and GST taken from each PDF, whatever its layout.

    documents.parse ×14
    14 / 14
  3. Match

    11 invoices match their purchase order and goods receipt. Three don't.

    erp.purchase_orderserp.receipts
    3 exceptions
  4. Investigate

    The agent works each exception on its own, choosing where to look.

    INV-209314.1% over PO. Found the supplier's 1 October price notice in the inbox; the rise is within contract terms. Approved, with the notice attached.Resolved
    INV-20952No matching PO. Same amount and lines as INV-20917, paid on 9 October. Rejected as a duplicate.Resolved
    INV-209442.7% over PO. No price notice or contract change on file. Drafted a query to the supplier for finance to send.To finance
    inbox.searchcontracts.lookuppayments.historyemail.draft
    2 resolved · 1 held
  5. Post

    Twelve bills drafted in the accounting system, coded and ready to pay.

    accounting.bills.create ×12
    12 drafted
  6. Log

    Every decision stored with the evidence behind it, plus cost and timing.

    trace.write
    Traced

What makes it an agent rather than a script: it decides how to handle each exception, gathers its own evidence, settles what it safely can, and hands the rest to a person with the work already done.

Agent-ready products

AI assistants and agents are becoming a new way customers use software. We design, build and host Model Context Protocol (MCP) servers that let assistants such as Claude and ChatGPT use your product safely: authentication, permissions, rate limits, logging and the tool design that makes agents use it well.

For software companies whose customers are starting to work through agents.

AI cost & performance

Model spend grows quietly with usage. We audit how your products and teams use language models, then cut cost and latency without losing quality: model routing, prompt and context design, caching, batching, and evaluation suites that prove nothing got worse.

For teams already running AI features in production.

Generative visual systems

Large, consistent visual and video libraries produced by engineered pipelines rather than one image at a time: every product, exercise, location or item in a catalogue rendered in one art-directed style, matched to your structured data, and checked before anything ships.

We build the pipeline (reference control, style systems, automated and human review, versioning) and can hand it to your team or run it for you as the catalogue grows.

The imagery on this site was produced the same way.

Automation

Most of the value is in the connections. We build event-driven workflows across CRMs, finance and ERP systems, inboxes, spreadsheets and internal tools, so work moves between systems without being re-keyed, and agents are triggered by the business events that matter.

Agent operations

Evaluation

Every agent ships with a test suite of real cases and measured accuracy, rerun on every change to prompts, models or tools.

Guardrails & permissions

Least-privilege access to systems, limits on what an agent can do unaided, and clear escalation paths.

Observability

Every run traced: inputs, decisions, tool calls, outputs, latency and cost, searchable and auditable.

Human in the loop

Approvals designed into the workflow where a decision carries risk, not bolted on afterwards.

Cost & performance

Model routing, caching and budgets so running costs stay predictable as volume grows.

How we deploy

Map the work

Find the processes where an agent will pay for itself, and define what good looks like.

Prototype on real data

A working agent against your actual cases within weeks, measured against the baseline.

Harden and launch

Evaluation, guardrails, monitoring and approvals in place before anything runs unsupervised.

Operate and improve

Ongoing monitoring, tuning and expansion to the next process.

Built on

Frontier and open models from the major and challenger labs, chosen per task; the Model Context Protocol for connecting tools; TypeScript and Python; and the same testing, review and infrastructure standards as the rest of our engineering.