Agentic AI systems
Agents that survive contact with production data — bounded, verified against your own system of record, and costed before a line of them is written.
The problem is arithmetic, not intelligence
An agent that answers five questions in a meeting room is not hard to build. The failure that matters shows up later and it is not a failure of reasoning: a step that succeeds 99.3% of the time — which is about the best per-step reliability anyone has independently measured — succeeds roughly 64% of the time when you chain sixty-four of them. At sixteen steps the same model finishes about 89% of the time.
Nothing is broken in that system. A good number was multiplied by itself until it stopped being good.
The second failure is quieter. Long loops are cheap to reason about and expensive to run, and nobody discovers the per-task cost until the invoice arrives. The third is worst: an agent with no verification path produces fluent, confident, unverifiable output, and the organisation has no way to tell the good runs from the bad ones.
"How would we know if it got it wrong?"
If a design cannot answer that question, the number of steps in it is irrelevant.
How we build them
Bound the loop, and make failure legible. A step budget means the agent either finishes inside it or reports that it could not. An agent that says “I could not establish this” is worth more than one that produces something plausible instead.
Take the deterministic work out of the model. Joins, graph traversals, filters and arithmetic all have correct answers, and a database produces them every time. When we costed a sixty-four-iteration design for legal case matching, the replacement pipeline pushed retrieval, graph expansion and verification entirely outside the loop — a single recursive query instead of a model deciding, repeatedly, how to walk a graph. Cutting depth from 64 to 16 bought about 25 percentage points of reliability and a fraction of the cost.
Separate the two error classes and treat them differently. A wasted retrieval costs tokens. A fabricated citation costs a client. Recoverable errors are a budget line. Fatal ones are driven to zero by construction — checking every identifier and every quoted span against the database on the way out — because verification is a property of the system and reliability is a property of the model.
What you get
An architecture document with a step budget and a cost per task. Tool definitions that reject nonsense arguments before execution. A verification layer between the model and the user. An evaluation set built before anything is tuned. And a written record of what we chose not to build, with the numbers that decided it.
When this is the wrong choice
- If a fixed pipeline answers the question, build the fixed pipeline. A large share of agent requirements are three function calls with a queue in front, and saying so is part of the work.
- If the output cannot be checked against anything authoritative, an agent will generate confident text at scale and nobody will be able to audit it.
- If the underlying data is wrong, an agent reaches the wrong answer faster and states it more persuasively. Fix the data first.
- If a human must own the decision — a filing, a diagnosis, a legal position — the agent’s job is to assemble evidence, not to conclude, and the interface should make that obvious.