Multi-agent systems and protocols
AI agents that do real work in your systems, under rules you set and people who supervise them
Chatbots answer questions; the work still waits for people
Most AI pilots stop at a chat window. The real cost sits in work that crosses systems: checking a request, looking up records, updating the ERP and telling the customer.
Giving an AI model open access to those systems is not an option for a regulated business. Every action must be allowed, logged and explainable, and a person must be able to step in.
What you get
Work done, not only answers
Agents that complete multi-step work across your systems around the clock, not only answer questions.
Every action accountable
Every tool an agent may use is defined, permissioned and logged, so actions can be audited.
People in charge
People supervise the agents and step in on exceptions, without approving every routine step.
Quality and cost under control
Quality, errors and token spend measured on every agent, with limits that stop runaway costs.
What we deliver
Multi-agent workflows with human-on-the-loop supervision
Agents that share a task, each with a narrow role, running around the clock. Under human-on-the-loop (HOTL) supervision, people watch the work and handle the exceptions.
MCP servers and enterprise tool use
Model Context Protocol (MCP) and WebMCP servers that give agents narrow, logged access to your APIs, databases and internal tools, with deterministic tool execution.
A2A interoperability and AP2 payments
Agent2Agent (A2A) lets your agents delegate work to agents in other systems. The Agent Payments Protocol (AP2) lets an agent pay under a cryptographic mandate you approve.
AgentOps
Telemetry, evaluation, hallucination guardrails and token cost control for every agent, so you can see what each one did, how well and at what cost.
How it works
Every tool call is deterministic and cryptographically auditable. AgentOps telemetry, evaluation and guardrails run on every agent, with human-on-the-loop supervision for the exceptions.
Choose the workflow
We pick a workflow with a clear business result, and agree which steps agents take, which people keep and where people must approve.
Define the tools
We expose the systems the agents need through MCP servers, each tool with its own permissions, inputs and logging.
Build and evaluate
We build the agents and test them against real cases, scoring accuracy, errors and cost before anyone relies on them.
Set up supervision
We give your team a view of what the agents are doing, alerts on exceptions and a way to pause or override any agent.
Release and operate
We release in stages, monitor quality and token spend, and extend the agents to new steps once the first ones prove themselves.
Technologies and standards
Agents work through open standards: MCP and WebMCP for tool access, A2A for delegation between agents and AP2 for payments. The ecosystems are built on Linux Foundation and FIDO standards.
- Frameworks
- FastAPI
- AI and machine learning
- LangChain
- LlamaIndex
- Methods and practices
- Human-on-the-loop (HOTL) supervision
- AgentOps telemetry and evaluation
- Hallucination guardrails
- Token cost control
How to start
Most clients start with the sovereign agentic proof of concept: a working multi-agent workflow connected to your internal tools in a private network.
What happens next
Agreeing a time
Our team emails you to agree a time for the session.
Pre-consulting session
A free 30-minute online pre-consulting session: we go through your goals, systems and constraints.
Readiness review
In a free 48-hour readiness review, a Principal Architect checks your project against the 6-point Definition of Ready and recommends the track that fits.
Choosing the track
Scope-ready projects receive a binding proposal within 3–5 business days. Concept-stage projects start in the Blueprint Studio. A live system that is failing, undocumented or needs integration starts with a diagnostic of 1–2 weeks. The Blueprint Studio and the diagnostic each have a fixed fee, quoted in the proposal.
AI-native product teams
Managed, cross-functional pods of senior engineers, equipped with agentic coding and QA harnesses.
Fixed-outcome projects and round-the-clock managed services
Fixed-scope, milestone-driven delivery with named delivery leads, and round-the-clock application managed services under SLA governance.
The quickest way in is the Sovereign agentic POC (MCP, A2A, GraphRAG) package, one of our rapid engagement packages.
Questions and answers
How is a multi-agent system different from a chatbot?
A chatbot answers questions. An agent takes actions: it reads records, calls your systems and completes steps of a process. A multi-agent system splits a larger task between agents with narrow roles, which are easier to test and control.
What are MCP, A2A and AP2?
They are open protocols for AI agents. The Model Context Protocol (MCP) connects an agent to tools and data. Agent2Agent (A2A) lets agents in different systems delegate work. The Agent Payments Protocol (AP2) lets an agent pay under a mandate you sign.
How do you stop an agent doing something it should not?
An agent can act only through the tools we define for it, each with its own permissions and limits. Every call is logged. Guardrails check inputs and outputs, and steps that carry risk wait for a person to approve them.
What does human-on-the-loop mean?
In human-in-the-loop (HITL) work, a person approves each step. In human-on-the-loop (HOTL) work, agents run the routine steps while people supervise, receive alerts and step in on exceptions. We agree with you which steps need which.
How do you know the agents are accurate enough?
We build an evaluation set from real cases with the expected results. Each agent is scored on it before release and after every change to prompts, models or tools. You agree the threshold, and we do not release below it.
How do you keep token costs under control?
AgentOps measures the tokens each agent and each step use. We set limits per task, cache repeated work and send routine steps to smaller models. Spend is reported, so a workflow that grows expensive is seen early.
Can the agents run without our data leaving our environment?
Yes. The agents, the tools and the models can run inside your own data centre or a private network on a regional sovereign cloud. Our sovereign LLM practice hosts the models there.
Who owns the agents you build?
You do. The code, prompts, tool definitions and evaluation sets we produce for you are yours once each milestone is paid. Commercial models remain under their own licence terms, which we list for you.
Start with a conversation, not a contract.
Tell us what you need to build, connect or secure. An architect goes through it with you, then recommends where to start.