The models we have available today are so capable and smart that the bottleneck is not the code generation itself. The AI doesn't understand the big picture.

David KayDavid KayInterim Chief Data Officer, City of Bellevue, Washington

Today's data and engineering teams are being asked to make agents reliable on systems whose rules were only ever in their staff's heads. An agent can read the schema, write the query, and return a number that satisfies every rule it can see while missing the policy that changed last quarter. Reliability now depends on what gets documented, who checks the output, and where a human stays in the sequence.

David Kay is Interim Chief Data Officer for the City of Bellevue, Washington. His mandate covers the city's data strategy and operating model, data and analytics engineering, and a governance program, and he partners with business intelligence teams on self-service analytics. He was previously Vice President of Data and Analytics Strategy at Tricon Residential, where he grew the data team into an organization spanning data engineering, business intelligence, data science, and governance. Kay also spent several years at Microsoft across its cloud and commercial businesses, and he's run his own consultancy advising executives on data strategy.

"The models we have available today are so capable and smart that the bottleneck is not the code generation itself. It's not writing SQL, it's not writing the software. The AI doesn't understand the big picture," Kay explains. He runs coding agents on his own side projects, and Bellevue is also now modernizing legacy systems now. Kay says AI readiness is one of the pillars guiding that work.

Context over code generation

Kay used to write prompts that told a model explicitly what to do, what to avoid, and what role to play, and he's stopped doing it. "You think about goal-oriented development, where you give it a goal, you define what good is or what done is, and then you let the agent figure it out," he explains. "The models are so good at reasoning and have so much underlying knowledge that the more guardrails you give it up front, the harder it will actually be for that agent to find the optimal path."

Instruction files hold the workflow for a particular task. A knowledge graph records how concepts relate, and an ontology does the same for business terms. Kay wants all of it available to the next session without passing the entire history forward each time, because a context window that grows too large starts dropping what's in its middle. "A semantic layer gives the business meaning to things like metrics or data fields or tables, beyond the metadata that would be in the database," he notes. "It's really helping ground the AI."

Most of what an agent needs to know about a dataset isn't stored in the database. A semantic layer keeps that context with the data, so the agent has something to ground its answers in, and that's where data intelligence starts.

Par BotesPar BotesVP, AI Infrastructure, Everpure

Par Botes, Vice President of AI Infrastructure at Everpure, works with enterprises preparing their data for AI agents. "Most of what an agent needs to know about a dataset isn't stored in the database," he explains. "What a revenue field counts, who owns it, and which version finance reports from usually live in someone's head. A semantic layer keeps that context with the data, so the agent has something to ground its answers in, and that's where data intelligence starts."

A company with clean data and a working semantic layer can still get the wrong answer out of a query. Kay says most haven't reached even that point, and his example grants it to them anyway. "Does the AI have the context to know that this time last year in the same quarter you had a pricing promotion?" he asks. "That's going to give you results that are mechanically correct, mathematically correct, but still missing the context to give you the good analysis that you would need."

From writing to reviewing

Coming into software from the data side, Kay doesn't write production code himself. The account he gives is secondhand, from engineers whose daily work has changed. "I've spoken to many colleagues and friends who are developers, and a lot of them, their jobs now today are just to review the code that the AI has written," he says. "Whether it's the agents themselves writing the code or the orchestration layer, they're very much just glorified PR approvers."

Orchestration is itself being handed to AI in some setups. In Kay's view, that reduces what a developer has to coordinate and compounds whatever goes wrong, because the developer sits further from the work. "The tax is shifting, because it used to be the tax of actually writing the code and debugging and reviewing it," he explains. "Now the tax, I think, is shifting away from the actual coding to more like the big picture software engineering perspective, where you have to think about architecture and then balancing. How do you manage the agents on top of that?"

A developer used to reach architecture and systems work after years on features. Kay says agents have taken the feature work, and how a junior developer now builds the judgment that came with it remains an open question. "Even a junior developer is being expected to think more higher level, more at a strategic level when it comes to developing software," he notes. "We're essentially asking them to do that with a flip of a switch. It's not a slow progression."

Agents across a team

Kay says some teams have hit the limits of how agent sessions work today. "How do you actually have more of a team based environment, where one developer can pick up a session from another developer, as opposed to that session being completely isolated on the one developer's local machine?" he asks. Databricks has since open-sourced Omnigent, a layer for combining and sharing agents between developers. Kay also points to Buzz, built by Jack Dorsey's Block, to put a team and its agents in the same channels.

Both tools are built for agents working across a team. "The data those agents read needs the same treatment," says Botes. "When every team keeps its own copy, two agents can give two different answers to the same question. One managed dataset, with one definition and one owner, lets every agent on a team work from the same numbers."

Anthropic replaced its internal source control with a system called Scion, after reaching about 30,000 code merges a day. Kay raised the example himself, as one end of a wide range. "That kind of sets a bar that a lot of companies are then not necessarily aiming to hit, but it shifts the perspective of what is the current AI native way of doing development," he says. "I think that we have to keep in mind that most companies are not full of PhDs being paid upwards of half a million dollars a year to do this stuff. They have to work within the confines of existing systems."

Checkpoints and institutional knowledge

Ideally, Kay wants knowledge work divided into stages, with a human checking the output of each one. He notes the form of that review was still being argued at a conference he had just attended, over whether the human sits inside the sequence or above it. "If you wait to the very end for a human in the loop, then if there is an issue, you have to do your root cause analysis through a much longer triage," Kay explains. "When the agent hands off its inaccurate artifact or product to the next agent, that's going to then compound the negatives of that agent."

When a policy is rewritten weeks after an agent has produced work under the previous version, Kay wants a way to trace those artifacts afterward. "If I look today at something that the AI produced, say a month ago, and that policy is now changed, what was the chain of custody essentially, or the provenance for that AI to have produced that artifact at that time?" he asks. "How do I assess a blast radius if something changes?"

Botes argues the traceability Kay wants has to live in the data itself. "Every dataset should carry its policy, where it came from, and why it exists," he says. "When a policy changes next month, you can find every copy and every output built under the old version and correct it automatically, without someone digging through systems to reconstruct what happened."

Documentation, structured files, semantic layers, and knowledge graphs all exist to hold what employees know, and getting it out of them is still unsolved. Employees asked to hand that knowledge over are often reluctant, and Kay traces it to what they hear from leadership, against a run of layoff headlines. He points to the companies now hiring staff back, a correction he attributes to cuts made too early. "Staffing changes should be a byproduct of your AI strategy. It's not in and of itself part of your AI strategy," he concludes. "If you are framing your strategy around cost savings by trimming staff before you actually have developed the ways of working that are AI native, then you're really putting the cart before the horse."