Essays
On making AI dependable
How agents behave in production, how to evaluate them, and the data, interfaces and controls they need inside businesses that have to answer for the result.
Series and essays
Series · 2 parts
Agent harness design
A series on the layer between the model and the world: the product concerns that decide whether an enterprise agent is reliable enough to ship.
- 01The cushions an agent loses in production
- 02The authority surface
-
When to use decision models
Which AI workflows are bounded decisions, how to turn a model's probabilities into thresholds from the cost of an error, and how to evaluate before automating.
-
The authority surface
Agent harness design, part II. Autonomy is not a dial. What an agent may do alone has to be decided action by action, on the cost of being wrong.
-
The record nobody agreed to keep
Why accountability for an agent's actions falls between the protocol, the schema and the regulation, and the record a firm will have to construct itself.
-
Encoding judgement
Agents rarely lack judgement because the model is weak. They lack it because the judgement was never written down anywhere the agent can reach.
-
A benchmark is not a control
Why a high leaderboard score cannot establish that an agent's output is acceptable in a regulated workflow, and how to read a benchmark for what it licenses.
-
An answer that runs is not an answer you can trust
A generated query that runs has proven only that it is valid. What a data-discovery harness must do to keep an analyst able to judge a result they can no longer inspect.
-
The cushions an agent loses in production
An agent that works on your laptop relies on supports that production removes one by one. Engineering them back is what reliability actually means.
-
Vibe Coding and the Boundary of the Firm
Cheap code has turned building into an option that reprices the whole make-or-buy decision. What should decide it are the costs that do not fall with it.
-
Agent-ready financial data: an architectural view
How a financial data provider makes its data ready for agents: the workflow archetypes, the human-in-the-loop bands, the delivery surfaces, and why the semantic layer is the product.
-
Ironies of AI automation
Bainbridge's 1983 framework applied to agents: why automating a task makes the remaining human more important and less able to do it, and why opacity makes it worse.
-
The catalog becomes the query interface
Why natural language changes the enterprise-data interface, and why the catalogue, semantic layer and governance model decide whether its answers can be trusted.
-
When the customer is an LLM
Why APIs increasingly serve non-human customers, and what that changes about documentation, predictability and API product management.