Devtools spent the week making agents first-class actors — with identity, governance, and deploy rights.

The week in devtools

The single most important move across devtools this week was the promotion of the AI agent from something a developer runs to something the platform recognizes as a principal in its own right. Identity vendors stopped treating an agent as a borrowed user session and started issuing it credentials, an audit trail, and scoped permissions; CI, infrastructure, and app-builder platforms opened surfaces the agent can call headlessly; and almost everyone who opened a surface shipped the governance to constrain it in the same window. The pattern that recurred product after product was reach first, then govern — expose an API or MCP server, then add the allowlist, policy check, or approval gate that makes it defensible.

The second thread was consolidation of the build-and-ship loop. Lovable and v0 by Vercel pushed AI app builders past the prototype objection by wiring in payments, connectors, and warehouse access, while infrastructure and CI platforms collapsed multi-step human workflows into single agent-callable operations. The net effect is a sector where the code is increasingly assumed to be machine-written, and the differentiated work has moved to identity, policy, review, and cost accounting around that code.

Leaders

Lovable had the densest week in the sector — five sparks against 22 improvements — running two tracks at once. Generated apps gained real integrations: built-in Paddle and Stripe billing wired in behind the scenes, plus a widening connector catalogue and a Shopify path. At the same time the workspace grew enterprise controls — audit logs, SSO, provisioning, and a publish-time security scan — answering the standing objection that AI-built apps are only prototypes from both the capability and the shippability ends.

Spacelift also posted five sparks, and its AI crossed the line from explaining infrastructure to provisioning it. Spacelift Intelligence can now deploy a module by name: Intent resolves it against the module registry, prompts for missing inputs, evaluates your Intent policies, and applies it as a managed resource — no module block and no pull request in between. The notable part is that the natural-language path runs through the same policy, state, and audit trail as everything else.

GitHub shipped four sparks that read as the accounting and control layer for AI-assisted development. MCP allowlists in enterprise managed settings make which tools an agent may reach an administered setting rather than a local choice; a Copilot impact dashboard now ties seat spend to pull-request output; and Copilot automations can fire on issue and PR comments. The reach-then-govern pattern is clearest here — the same fortnight that widened what agents can do also shipped the telemetry to price it.

Auth0 and WorkOS converged on the same idea from the identity side: an agent should be a registrable principal, not a shared client or a human's session. Auth0 moved Agents as Principal and a Token Vault Privileged Worker into Early Access, giving background agents their own identifiers and credentials, and put Enterprise Connect into beta to be adoptable without a migration. WorkOS shipped Agent Registration — agents sign into AuthKit as their own identity via the auth.md protocol — alongside a Pipes Token Proxy that holds third-party credentials so the calling app never touches the token.

Wildcards

Windsurf is the off-pattern story of an editor being absorbed into its acquirer. Its own feature list has thinned; what ships now is Devin arriving one surface at a time. This week Devin Review and Quick Review reached every IDE user at no extra cost, framed explicitly around reviewing AI-generated code — the pitch is no longer that the agent writes the code but that the editor checks what the agent wrote, and it lands by bundling rather than upselling.

webpack went the opposite direction from every AI headline and quietly absorbed its own loader ecosystem. Version 5.109.0 defaults the TypeScript, CSS, HTML, and async-WebAssembly experiments to "auto" — on Node 22.6+, TypeScript compiles with no loader registered — and adds native build-progress reporting and automatic ESM resource hints. It is webpack conceding that config assembly was the problem it should solve itself.

Themes that compounded

• Agent-as-principal identity: Auth0 and WorkOS both issued agents their own credentials and token custody instead of shared accounts.
• Reach-then-govern: GitHub MCP allowlists, Spacelift Intent policies, and Cloudflare's agent onboarding all paired a new surface with its control.
• Headless, agent-callable operations: Buildkite collapsed failed-build triage into one MCP call and added job-promise signalling, while BigQuery took its MCP server and in-SQL AI functions to GA.
• App builders past the prototype line: Lovable added payments and v0 by Vercel added Figma inspection, Snowflake, and a Platform API so a person no longer has to be in the loop.
• Removing the syntax barrier: Honeybadger made its query language optional via natural-language search, and Windmill turned AI sessions on by default while running dbt projects unmodified.

Watch this week

Watch whether the agent-identity surfaces start moving from Early Access and beta toward GA. Auth0's Agents as Principal and Privileged Worker are still Early Access, and WorkOS's Agent Registration and Pipes Token Proxy are new enough that their convergence — a registered identity scoping which third-party calls the proxy will make — is the obvious next step both roadmaps point at. On the governance side, GitHub's impact dashboard currently ties spend only to pull-request output while its usage API already carries per-agent activity, so per-agent cost breakdowns are the field to watch. And with Spacelift now deploying single modules by name, whether Intent extends to multi-stack changes with policy checks in the loop will show how far the deploy-by-language pattern can stretch.