AI Agent Development Services
LLM agents that use tools, call your APIs and complete multi-step work, with the guardrails, observability and human checkpoints to deploy them safely.
- Typical duration
- 10–16 weeks
- Investment
- from €75,000
- Team
- Two engineers plus a fractional architect
Short answer
Rubra builds LLM agent systems that take actions rather than just answering questions: tool and API integration, multi-step planning, state management, human approval checkpoints, and full execution tracing. I am deliberately conservative about where agents belong, because most tasks marketed as agentic are better served by a deterministic workflow.
Last reviewed
Agents are the most oversold and most under-engineered category in applied AI right now. The demos are extraordinary. The production deployments are, mostly, narrow, carefully bounded and much less exciting than the demos. That is precisely why they work.
My bias: fewer agents, better bounded
Before building an agent I ask whether the task actually needs one. Most do not. A large share of what gets pitched as agentic automation is a workflow with a language model at two or three decision points, and building it that way makes it cheaper to run, faster, and possible to debug at 3am.
When the path really is unknowable in advance, an agent is the right tool. I then spend my effort on the things that determine whether it survives contact with reality.
What determines whether an agent works
Tool design, not prompt design. An agent is only as good as the operations available to it. Tools with narrow, typed inputs and informative error messages produce reliable agents; tools that accept free-form strings and fail silently produce agents that loop. Most of my engineering time goes here.
Bounded autonomy. Every agent runs inside a budget covering steps, wall-clock time and spend, and every irreversible action passes through an approval gate. The gate is a product decision as much as a safety one: users trust a system more when they can see what it is about to do.
Observability from the first commit. Every run is fully traced and replayable: what the model saw, what it decided, which tool it called, what came back. Without this, debugging an agent is archaeology.
Failure containment. The failure modes get enumerated during design: wrong tool, right tool with wrong arguments, plausible-but-wrong intermediate conclusion, infinite loop. Each one gets its own containment. The question is never whether the model will get something wrong, only what happens when it does.
Where I have seen agents deliver
Support triage that reads a ticket, checks three internal systems and either resolves it or routes it with a written summary. Vendor and contract review that pulls a document apart against a checklist and flags the clauses a human needs to read. Data reconciliation across systems that disagree, where the agent investigates the discrepancy and proposes the correction for approval.
The pattern in all of them: bounded scope, reversible actions, a human on the consequential decision, and a measurable completion rate.
What you get
- Tool and API integration layer with typed, validated schemas
- Planning and orchestration logic with bounded retries
- Permission model and human-in-the-loop approval checkpoints
- Full execution tracing and replay for every run
- Failure-mode analysis and containment design
- Cost and latency budgets enforced at runtime
Outcomes
- Multi-step work completed end to end, with a human checkpoint where it matters
- Every action traced, replayable and reversible
- Clear boundaries on what the agent may do without approval
Frequently asked questions
When is an agent the right choice over a fixed workflow?
An agent earns its complexity when the sequence of steps cannot be known in advance, because the path depends on what earlier steps discover. Triage across heterogeneous inputs, investigation and research tasks, and exception handling in an otherwise deterministic process are good fits. If you can draw the flowchart, build the flowchart: it will be cheaper, faster, more reliable and far easier to debug.
How do you keep an agent from doing something harmful?
By constraining what it can reach rather than by asking it nicely. Tools are scoped to the minimum permission needed, destructive operations require explicit human approval, every action is logged before it executes, and spend and step budgets terminate runaway loops. The design assumption is that the model will eventually attempt something wrong, so the containment has to hold without depending on the model behaving.
What does agent reliability actually look like in production?
Reliability comes from narrow scope and good tools, not from a better prompt. A well-scoped agent with five well-designed tools and clear success criteria is far more dependable than a broad one with thirty. I measure task completion rate, intervention rate and cost per completed task, and I tighten scope until those numbers are acceptable rather than adding instructions until the demo passes.
Which agent framework do you use?
Whatever fits, and often none. Frameworks are useful for tracing and orchestration scaffolding, but they also hide control flow you will eventually need to reason about during an incident. For most production systems I write the orchestration explicitly and use libraries for the parts that are tedious. I am comfortable working inside whichever framework you have already adopted.
Get a straight answer on your AI roadmap
A 30-minute call with the engineer who would do the work, not a salesperson. You will get an honest read on what is worth building, what is not, and roughly what it costs.
No NDA needed to talk. EU and UK hours in full, with afternoons overlapping US Eastern and Central.