How we work.
Method, patterns, and operating model — codified from production systems and re-applied with discipline.
Sprint, Build, Retainer.
The engagement arc is deliberate. A Sprint is the loss-leader — fixed fee, two to three weeks, a maturity assessment and roadmap that pays for itself if the client never engages further. A serious buyer can act on the Sprint output alone.
A Build follows when the Sprint identifies leverage. We deploy the Agent Operating System against the highest-value bottleneck first, in four to twelve weeks, milestone-by-milestone. We do not run open-ended engagements. We do not bill for discovery a second time.
A Retainer follows the Build when the system needs care. Quarterly reviews benchmark performance against the McKinsey-derived envelope — 20% EBITDA uplift, 3:1 ROI, 1–2 year breakeven — and compensation links to those numbers. If the system stops delivering against them, the retainer ends.
Built on Claude,
end to end.
The Agent Operating System is designed around Anthropic's Claude — not bolted onto it. That means model selection is a deliberate engineering decision, the developer platform is used as intended, and governance is built in rather than retrofitted. We work in Claude every day, on our own operations before any client's.
The point is not enthusiasm for a vendor. It is depth in one ecosystem, held long enough to know where it is strong, where it is not, and how to build systems that age well as the platform moves.
Opus, Sonnet, and Haiku matched to the task — reasoning depth against latency and cost, decided per workflow rather than defaulted.
The API used as designed — prompt caching, batch processing, streaming, and structured tool use, instrumented for cost and quality.
Capabilities packaged as versioned Skills and delegated to subagents, so the system stays legible and each part can be tested in isolation.
Enterprise systems exposed to Claude through governed MCP servers — the single seam where scope, identity, and audit are enforced.
Used in delivery itself — the firm ships its own build work through it, which is why estimates reflect what one architect can actually carry.
Behaviour measured against written evals before deployment; safety, refusal, and review surfaces designed in, not appended.
Anthropic builds the model and the platform. We build the operating system, the governance, and the accountability around them — and we never imply otherwise.
Five patterns. Each one battle-tested before it ships.
The library is the firm's leverage. Each pattern is documented, instrumented, and reused. New engagements pay for novelty only where novelty is warranted.
The Gateway sits between agents and the systems-of-record. It enforces scope at the tool level, attaches caller identity, logs every invocation, and rate-limits per-agent and per-tenant. Without it, agents accumulate trust the organisation cannot reason about.
In regulated contexts, the Gateway is the audit surface. Compliance reviews look at it, not at the agent's prompt.
// MCP tool definition — registered on the gateway { name: 'list_invoices', description: 'List unpaid invoices for a tenant.', scopes: ['finance.read'], rate_limit: '60/min', audit: 'full', handler: async (ctx, args) => { /* … */ } }
The Portal Shell is where humans stay in the loop. Agents do the drafting; humans do the review and the irreversible commit. The Shell makes that boundary structural — not a checkbox.
Every panel renders agent output adjacent to its inputs and audit trail. Approvals carry the reviewer's identity into the system-of-record.
Most clients have ten to twenty services that an agent benefits from reading or writing — CRM, ERP, document store, calendar, billing, ticketing. The template gives each one a consistent MCP surface in under a day of work each.
Naming, scoping, error handling, and audit all share one shape. Agents learn one pattern, not twenty.
Prompts, skills, and subagent definitions are code — versioned, reviewed, and shipped with the same rigour as the application around them. We never paste prompts into a console and call it deployed.
Configuration travels with the engagement: a client receives the full kit, not a black box.
Chat is the wrong default for most enterprise workflows — but the right default for ad-hoc questions, drafting, and explanation. Portal Chat lives inside the Shell, sees the same state the user sees, and routes to the same governed tool surface.
It is not a separate product. It is the conversational facet of the Portal.
The Agent OS, in five layers.
Our reference architecture is informed by publicly-discussed AI-maturity frameworks — DAIN among them — but the five-layer model below is our own. We name what we built and credit what we built on; we don't claim either as the other.
The Agent Operating System — our five-layer reference architecture. "Agent OS" describes our implementation, not a category we claim to own.
The reference architecture deploys as CLTF — the firm's codified platform, currently in production at the lab and on selected client engagements. Client-integrated deployments adapt the platform into the client's own infrastructure; hosted engagements run on instances we operate. The platform is the firm's leverage; the engagement is the work.
Agent-native, by construction.
The principle is short: control planes get MCPs, not UIs. Anywhere a tool exists primarily to coordinate work, we expose it as an MCP server and let agents drive it. Anywhere a tool exists for composition or review, the human surface stays.
Internally, the firm runs the same shape: a TobDog scheduler, an EA agent, a business-development agent, a delivery agent, and a CFO agent. They run the practice between client engagements. That is what makes a one-architect consultancy possible without theatre.
Start with a Sprint.
Two to three weeks, R60k–R80k or £5k–£8k fixed fee, delivers a maturity assessment and a phased roadmap.