AI agent development in London
Agents that do work rather than describe it — tool-using systems built on Claude and GPT, with the evaluation harness and audit trail that separate a production agent from an impressive demo.
The short answer
An AI agent is a language model given tools and a goal, which decides for itself which tools to call and in what order. Build one when the number of valid paths through a task is too large to enumerate; build a plain workflow when the steps are known. A single-purpose production agent typically takes four to eight weeks, of which the evaluation harness is often a third — and is not optional.
On this page
What an AI agent actually is
The word has been stretched to mean almost anything, so here is the working definition we build against: an agent is a language model given a set of tools and a goal, which decides for itself which tools to call, in what order, and when it is finished.
Two distinctions do most of the useful work. The difference from a chatbot is action — a chatbot returns text, an agent changes something: it updates a record, sends a message, files a ticket, runs a query. The difference from a workflow is control — in a workflow you decide the sequence in advance, in an agent the model decides at runtime. That second distinction is where projects are won and lost, because handing control to the model buys flexibility and costs predictability, and plenty of teams pay that price without needing what it buys.
When an agent is the wrong answer
We turn down more agent projects than we take, usually in favour of something simpler that will work better. The honest test is whether you can enumerate the valid paths through the task.
Build an agent
Open-ended research, triage that spans many systems, debugging, anything where the next step genuinely depends on what the last step returned.
Build a workflow
Known steps in a known order. A deterministic pipeline with a model in one or two slots is cheaper, faster, easier to test and far easier to debug at 3am.
Most agent projects that fail should have been workflows. Teams reach for the agent first because it is the more exciting artefact, then discover they have bought non-determinism they did not need and cannot now test. If your task decomposes cleanly, our AI automation practice will get you there for less money and more reliability.
Agents we build
Internal operations agents
The highest-return category and the least glamorous. An agent that reads an inbound request, pulls the relevant context from your CRM and knowledge base, drafts the response and files it for approval — turning a queue that eats a day a week into a review task measured in minutes.
Research and analysis agents
Given a question, the agent plans its own search, gathers sources, reconciles conflicts and returns an answer with citations you can check. The value is not the summary; it is that every claim traces back to a document, which is what makes the output usable in a decision.
Customer-facing agents
The hardest bar, because an unreviewed message to a customer is irreversible. These need tight tool scoping, a much stronger eval suite, explicit escalation paths and — in most cases we recommend — a human approval step for anything outside a narrow confidence band.
Multi-agent systems
Occasionally the right shape: a coordinator delegating to specialists with separate tools and separate context. Genuinely useful for large parallelisable work, and genuinely more expensive to build, test and reason about. We recommend it when the parallelism is real, not because the architecture diagram looks impressive.
How we build them
We are vendor-neutral and deliberately keep the model swappable — the field moves fast enough that hard-coding a provider is a liability. In practice most production agents we ship run on Anthropic Claude or OpenAI GPT for tool use, with open-source models where cost, latency or data residency argue for it.
What matters more than the model:
- Tool design. Agents fail at tool boundaries far more often than at reasoning. Tools need narrow, well-described interfaces and error messages written for a model to act on, not a human to read.
- Context strategy. What the agent can see, when, and how it is retrieved. Long-running agents live or die on context management — often via a retrieval layer rather than an ever-growing prompt.
- Evaluation harness. A fixed set of realistic tasks, graded automatically, run on every change. Task completion, tool-call correctness, cost and latency per task, and the rate of looping or unsafe behaviour.
- Observability. Every tool call logged with inputs and outputs, so that when an agent does something surprising in month three you can reconstruct exactly why.
Keeping agents safe in production
Prompt instructions are not a control. An agent told not to do something will still occasionally do it, and "we asked it nicely in the system prompt" is not a defensible answer to an auditor. Real control comes in three layers:
- Scope the tools. The most reliable safeguard is unreachability — an agent without delete permission cannot delete, regardless of what it decides.
- Gate the irreversible. Anything outward-facing or destructive gets a human approval step. This is a design decision, not a limitation, and it is what makes agents deployable in regulated environments.
- Log everything. Full tool-call traces with inputs, so any action can be reconstructed and explained after the fact.
Agents that read untrusted content — web pages, inbound email, uploaded documents — need one more consideration: that content can contain instructions aimed at the model. We design those agents so retrieved content is treated as data rather than direction, and so anything consequential still passes a human.
How agents fail — and what it costs
Three failure modes account for most of the trouble, and all three are cost as well as correctness problems.
Looping. An agent that cannot achieve its goal will often keep trying, burning tokens on each attempt. Hard iteration caps and per-task cost ceilings are standard in everything we ship.
Silent wrongness. The agent completes confidently and incorrectly. This is why the eval harness is not optional — it is the only mechanism that catches degradation before your users do.
Cost drift. An agent that was economical in testing meets real-world inputs three times longer than your fixtures. We forecast per-task cost during discovery and instrument it in production, because AI projects are far more often killed by an unexpected invoice than by poor quality.
Where agents fit with the rest of the work
Agent development sits inside our broader AI development practice. If you are not yet sure whether an agent is the right answer, a short consultancy engagement will tell you in one to two weeks — and quite often the answer is a simpler automation that ships sooner. Full costs for every engagement shape are on our pricing page.
Frequently asked questions
What is an AI agent?
An AI agent is a language model given tools and a goal, which decides for itself which tools to call and in what order until the goal is met. The distinction from a chatbot is action: a chatbot returns text, an agent changes something. The distinction from a workflow is control: in a workflow you decide the sequence, in an agent the model does.
When should you build an agent instead of a fixed workflow?
Build an agent when the number of valid paths through a task is too large to enumerate — open-ended research, triage across many systems, debugging. Build a fixed workflow when the steps are known, because a deterministic pipeline is cheaper, faster and easier to test. Most projects that fail as agents should have been workflows.
How much does AI agent development cost?
A single-purpose agent with two or three tools typically takes four to eight weeks to reach production. Multi-agent systems, or agents in regulated environments where every action needs an audit trail, take considerably longer and cost proportionally more. The evaluation harness is often a third of the build and is not optional. See how we price.
How do you stop an AI agent doing something harmful?
Three layers. Scope the tools so the destructive action is not reachable — an agent without a delete permission cannot delete. Put a human approval step in front of anything irreversible or outward-facing. And log every tool call with its inputs so you can reconstruct exactly what happened. Prompt instructions alone are not a control; they are a suggestion.
Which model is best for building agents?
Claude and GPT are both strong at tool use and are what most production agents run on today. The right choice depends on your tool-calling patterns, latency budget, context requirements and governance constraints rather than benchmark position — and we design agents so the model is swappable, because the field moves too fast to hard-code a vendor.
How do you test an AI agent?
With an evaluation harness: a fixed set of realistic tasks, graded automatically, run on every change. You measure task completion, tool-call correctness, cost and latency per task, and the rate of unsafe or looping behaviour. Without this you cannot tell whether a prompt change improved the agent or just moved the failures somewhere you were not looking.
Talk to us about an agent
Tell us the task you want an agent to own and we will tell you honestly whether it should be an agent at all. Or email hello@brashpixels.com.