Devtools spend the week turning agents into governed, billable platform primitives.
The week in devtools
The dominant move this week was not a new agent capability but the plumbing around it. Across the largest products, the cloud coding agent stopped being a feature you toggle and became a primitive you call, meter, and govern. GitHub exposed Copilot cloud agent tasks as a REST API, Cursor shipped custom tools and nested subagents through its SDK, Depot took its CI API and CLI to GA, and Auth0 brought machine-to-machine access for third-party apps to GA framed explicitly around AI agents. The common shape is an agent that lives behind an API and an account, not behind a button in an editor.
The second pattern is that finance and identity teams are now in the loop by design. GitHub flipped Copilot to usage-based billing for all users in the same window it shipped the agent API; Vercel moved function invocations to per-unit billing; Cursor shipped Organizations with per-team budgets. When a platform makes agents programmable and high-volume, it also has to make them meterable and capped, and the changelogs show both halves landing together.
Leaders
GitHub had the densest week, with four sparks against nineteen improvements. The Agent tasks REST API turns the Copilot cloud agent into an endpoint other tools can drive, and it landed alongside usage-based billing going live for everyone and the Copilot SDK reaching GA. Read together, these convert Copilot from an IDE add-on into a metered, programmable platform that enterprises deploy rather than toggle.
Vercel kept stacking models onto AI Gateway while hardening the layer beneath it. The standout is Docker support inside Vercel Sandbox: an agent can now build images, install system packages, and run container workloads without touching the host, pushing Sandbox toward a remote developer workstation for agent-as-developer use. Per-unit function billing and a Terms rewrite covering AI-initiated account actions point the same direction.
Depot made its platform pivot official. The Depot CI API and CLI hit GA off a shared OpenAPI contract, exposing every dashboard action to terminals, scripts, and agents. Paired with a new JUnit test-results product and the Sherlock assistant reaching into CI workflows, Depot is becoming a CI system agents operate, and its test analytics reach into GitHub Actions too as a migration wedge.
Cursor extended its agent platform with SDK custom tools, auto-review of local tool calls, and arbitrarily nested subagents, the clearest sign it wants agents running in production and CI rather than only inside the editor. Organizations reached GA in the same window, adding multi-team governance and budgets, so the programmable and governable halves advanced together.
Auth0 is assembling the identity layer for non-human clients. M2M support for strict third-party applications reached GA with organization-scoped tokens, openly framed around AI agents and partner backends with no user in the loop, while Custom Token Exchange added RFC 8693 delegated authorization in Early Access. The throughline: agents act for users, and every token records who-on-behalf-of-whom.
Wildcards
Kubernetes ran counter to the agent-platform rush by spending v1.36 clearing inherited debt. The official Dashboard was archived in favor of Headlamp, Service externalIPs was deprecated over a six-year-old CVE class, and etcd 3.7 entered beta. The AI angle here is upstream scheduling primitives for batch and ML workloads, not a packaged agent product.
Elasticsearch spent the week almost entirely in security-hardening mode under the elastic advisory feed: a synchronized drop of Kibana CVE fixes (SSRF, DoS, privilege escalation) across four branches, with Rally 2.13.0 the only feature-bearing release. It is the inverse of the sector's expansion story, a product being patched rather than extended.
Themes that compounded
- Agents became API primitives: GitHub, Cursor, and Depot all exposed agent or platform actions programmatically in the same week.
- Billing got coupled to usage: GitHub usage-based Copilot billing and Vercel per-unit function billing both prepare for variable, compute-heavy agent traffic.
- Identity stretched to non-human clients: Auth0 third-party M2M and delegated token exchange make agent access a first-class, auditable surface.
- Enterprise governance landed alongside capability: Cursor Organizations and GitHub budget and cost-center work pair new power with the controls to cap it.
- Test and CI intelligence moved up the stack: Depot added test analytics and AI failure diagnosis on top of its build-speed base.
Watch this week
Watch whether the agent APIs that shipped in preview start moving toward GA and gaining the governance that the same vendors shipped this week: GitHub's Agent tasks REST API alongside its now-GA billing and cost-center controls, Cursor's SDK under the new Organizations budgets, and Auth0's Early Access delegated token exchange behind its now-GA third-party M2M. The clearest near-term tell is repricing, with both GitHub and Vercel having coupled cost to usage in this window, so expect the next changelogs to refine those meters as real agent traffic arrives.