InPublishing: Peter Barr-Watson on the agentic shift in publishing
Peter Barr-Watson, chief technology officer at The Independent, published this column in InPublishing’s March 2026 issue. It argues that the useful frame for AI in publishing has shifted again: from chatbot (2024) to copilot (2025) to colleague (2026). The practical difference is not in what the technology can do but in how teams structure their relationship to it.
The agentic harness
The central concept Barr-Watson introduces is what he calls the “agentic harness” — treating AI not as a tool used on demand but as an active participant in a workflow with defined responsibilities and checkpoints. His worked example splits a finance report across three agents: one retrieves market data, one runs calculations in code, and one checks for anomalies against historical baselines. Validation gates between each step allow the process to halt automatically and surface a problem for a human to review.
The point of the example is not that this configuration replaces a financial journalist — it doesn’t — but that the editor’s role changes. Instead of doing each step, the editor is designing the workflow, reviewing the gates, and making calls on the edge cases the system flags. The column argues this is a better use of skilled editorial judgment than tasks that agents handle well.
Composable architecture over vendor suites
Barr-Watson makes a specific argument against buying all-in-one AI content platforms, which he describes as difficult to adapt when the underlying models improve or when editorial needs shift. He advocates for composable architectures built in-house from smaller, purpose-specific tools. The tradeoff is engineering cost upfront for flexibility over time, and he frames it as a decision about which dependencies a publisher wants to have.
Audience applications
On the reader side, he proposes using agents to personalize individual newsletters based on each reader’s reading history — not to generate content but to curate it from what the newsroom has already published. He connects this to reducing dependence on search traffic, since a newsletter curated for a reader’s specific interests gives them a reason to subscribe directly rather than arrive through a search engine.
Who it is useful for
The column is written for editors and digital publishing leaders making decisions about AI tooling, not for engineers. It assumes no technical background but does expect familiarity with how newsrooms work. It is most useful as a framing document — it won’t tell you what software to use, but it will give you a vocabulary for the conversation about what kind of colleague you want AI to be.