AI agent development
Most AI pilots stall because a demo that answers questions is not a system that does work. We build agents that take actions inside your tools — with the permissions, evaluation and monitoring needed to let them run unsupervised.
Why most agent projects never reach production
A chat prototype takes an afternoon. Getting the same thing to run reliably against real customer data, with real consequences for being wrong, is a different engineering problem — and it is where most internal AI initiatives quietly stop.
The gap is rarely the model. It is everything around it: which tools the agent may call, what happens when it calls them wrongly, how you know whether a change made it better or worse, who gets alerted when it fails, and how you prove to your own risk team that it is safe to let it act.
That scaffolding is unglamorous and it is most of the work. It is also the part we have standardised, which is why our agents ship rather than demo.
What we actually build
Tool-using agents
Agents with scoped access to your real systems — CRM, ticketing, database, internal APIs — through typed function calls with validation on every argument. The agent proposes; the tool layer enforces what is permissible.
Multi-step workflows with state
Long-running processes that survive restarts, retry sensibly on failure, and can be paused for human approval at defined checkpoints. Built as explicit state machines rather than a prompt hoping for the best.
Human-in-the-loop approval
The agent drafts an action and a person confirms it, until the logs justify loosening that. Approval thresholds by action type and value, so low-risk work runs unattended while consequential steps still get a human.
Evaluation harness
A test suite of real cases with expected outcomes, run on every prompt or model change. Without this you cannot tell an improvement from a regression, and you are tuning blind.
Guardrails and cost control
Input and output filtering, rate and spend caps per user and per workflow, and hard stops on destructive operations. Token spend tracked per feature so cost is attributable rather than a single alarming invoice.
Observability
Every run traced end to end — inputs, tool calls, intermediate reasoning, final output, latency and cost. When something goes wrong you can see exactly which step did it instead of guessing.
Typical stack
Chosen per project, not by habit. If your team already runs something that works, we use it.
- Models
- Anthropic Claude
- OpenAI
- Google Gemini
- open-weight models where data residency requires it
- Orchestration
- LangGraph
- Model Context Protocol (MCP)
- custom state machines
- queue-backed workers
- Retrieval & memory
- pgvector
- Pinecone
- Qdrant
- hybrid keyword + vector search
- Evaluation & monitoring
- LangSmith
- Langfuse
- custom eval suites
- OpenTelemetry tracing
- Runtime
- Python
- TypeScript
- Docker
- AWS / GCP / Azure
- your existing cloud
How an agent project runs
We start with one workflow, not a platform. A single agent doing one job well produces a decision about whether to expand — a general-purpose assistant produces a demo nobody owns.
- 01
Scoping
One to two weeks. We map the workflow, identify which steps are genuinely automatable, and write a technical scope with a fixed price for the build. You keep the document either way.
- 02
Pilot
Four to eight weeks to a working agent handling real cases in a controlled subset — usually with human approval on every action while the evaluation data accumulates.
- 03
Hardening
Approval thresholds loosened where the data supports it, monitoring and alerting wired to your on-call, runbooks written, cost controls tuned.
- 04
Run
We operate it, or hand it over with documentation your team can act on. Most clients start with us running it and take it in-house once it is boring.
Common questions
How is an AI agent different from a chatbot?
A chatbot answers. An agent acts — it calls your systems, changes records, triggers workflows, and decides what step to take next. That difference is why an agent needs permissions, approval gates, evaluation and monitoring, and a chatbot mostly does not.
What does an AI agent project cost?
Scoping is a fixed fee in the low thousands and produces a technical document you keep whether or not you continue. A first production pilot typically lands in the range of a few weeks of engineering. We give a fixed price for the build before you commit, so the number is agreed rather than discovered.
Will our data be used to train models?
No. We use enterprise API tiers where the provider contractually excludes your data from training, and for regulated or sensitive workloads we deploy open-weight models inside your own cloud so nothing leaves your perimeter. Which route suits you is settled during scoping.
What if the agent makes a mistake?
That is a design assumption, not an edge case. Consequential actions run behind human approval until the evaluation data justifies loosening it, destructive operations have hard stops, and every run is traced so a failure can be diagnosed rather than guessed at.
Can you work with our existing systems?
Yes — that is the normal case. Agents connect to what you already run through APIs, database access, or existing integration layers. Where a system has no API we work through the same interfaces your staff use, though we will tell you honestly when that is fragile.
How long until we see something working?
A working agent handling real cases in four to eight weeks from the end of scoping, for a single well-defined workflow. Anyone promising a production agent in a fortnight is describing a demo.
Start with a scoping call.
Thirty minutes, no obligation. If we are not the right fit we will tell you on the call rather than after a proposal.
Related services
AI workflow automation
The processes that resisted automation for twenty years were the ones needing judgement: reading a document, d…
LLM & RAG integration
Getting a model to answer questions about your own content is easy to prototype and hard to trust. The differe…
AI readiness audit
A short, fixed-fee engagement that answers three questions: where AI would actually pay off in your business, …