AI Agent
Consultancy
AI pilots don’t fail from bad models. They fail from unclear authority.
Give it one clear job
Limit what it can use
Test the results
You get an agent your team can run and support.
Authority, not model quality
You need to specify the agent’s data access, task boundaries, and who owns operations once it’s running. We can help you get your project across the line with practical implementation.
Is your team considering AI agents, but struggling to find a place to start? We can help you get your head around what’s possible. Book a call to discuss how to move forward.
What the Agent Review covers
Book a fixed-price Agent Review. We’ll dig into what’s involved in building AI agents into your systems, including:
A technical assessment of the systems your agent will talk to, including a discussion of permissions and capabilities
A risk assessment: what could go wrong, what’s immediately worrying, what needs addressing
A practical assessment of where there’s clear, immediate business value in agent-based automation
Recommendations: what needs building, how we’d go about it, and what it might cost.
EU AI Act applicability and requirements, if you’re building agents that will be used in the EU.
What we mean by an agent
What do we mean when we say agents? Anthropic make a useful distinction between workflows and agents: workflows are systems where LLMs and other tools are orchestrated through hardcoded paths. Agents are those where the LLM directs its own journey and tool use, operating autonomously towards a goal.
The line isn’t always clear, but there are lots of tasks that are more workflowy: they have clear steps and predictable behaviour. Building them as agents means your agent will struggle to perform a three-step process. Agents can also be unpredictable, calling the wrong tools and looping. They’re powerful, but should be used only where necessary.
The questions we settle first
Agentic AI is fashionable, but agents need boundaries in order to be reliable and useful. We help teams understand what’s feasible and useful before committing to a project.
- What’s the blast radius? Can it access your entire database? Can it email your entire company? Can it give refunds? How much could it refund? Who approves refunds? How much autonomy should it have, and where should it ask humans? We write this down and build accordingly.
- What’s the input? Are you pointing it at emails, websites, other untrusted sources? Agents can easily be misled by prompt injections embedded in their inputs. Wherever possible, we limit access to minimise impact.
- Where’s the entrypoint? Agents are complex. We insist on one clear entrypoint so it’s obvious where the system is being called from. This makes it easier to secure and understand.
- Where’s the off-switch? Autonomous processes need to be turned off. Depending on the task, this can be a killswitch in your infrastructure, or simply refusing to serve requests from a particular domain. Either way, someone should be responsible for pulling the plug.
- How will you know it’s working? We create a suite of test prompts, a benchmark, to verify that your agent is doing the right thing, and run them each time you make a change. These evaluations give confidence in making changes.
- Who approves spending? Agents should be restricted to specific, tightly-scoped tasks. Any action which incurs a cost, or communicates with third parties, should be approved by a human. Agents are useful, but we’re careful to automate tasks that are predictable and reversible.
- Can you see what happened? Agents should be observable. We log prompts, responses, tool calls and costs for every interaction so you can replay them later. Otherwise, you won’t know it’s looping until the credit card statement arrives.
- Can it do SQL? Can it execute commands? Agents work best with small, tightly-scoped tools which can be understood and reasoned about. We avoid exposing powerful primitives like shell or database access.
How we work
Book a call to find out how AI agents might work in your stack. We’ll dig into your requirements and advise on the cheapest way to solve your problems, and it may not be an agent. Reviews are fixed-price. If you’re happy to continue working together, we’ll scope a fixed-price project for the first phase of build. We’re pragmatic and experienced Laravel devs, happy to use libraries where it makes sense, and happy to discuss agent-specific tech in more depth. You own all code and repos. We’ll leave a dashboard for monitoring, a roadmap for extending it, and a runbook for when things go wrong. Ongoing support and development is available on a monthly retainer if you want it, but you’ll have everything you need to continue without us.
Get in touch
If you have an agent-based AI project that’s not quite there, get in touch: we can probably help. If you’ve got something you want to automate and are not sure whether agents might be helpful, we can advise on that too. It’s often cheaper to look at requirements before building than afterwards.