Development's fastest movers are building for the agent as caller, and the governance layer to keep it accountable.

The week in development

The clearest directional move this week is products treating the AI agent as the caller they build for, not the human. Of 151 products that shipped, 80 landed at least one spark, and the fastest-moving converged on one idea from different directions: make the control plane fast, scriptable, and governable enough to sit inside an agent's loop. Appwrite rewrote its CLI in Go for ~10ms startup because something now calls it in a loop; GitHub co-published a plugin standard that runs across competing agent clients; Rollbar closed the loop from production error to opened pull request. The pattern is capability expansion pointed squarely at machine callers.

The second, quieter half is governance and verification of whatever the agent produces. HashiCorp stated the thesis outright — the control plane's job is to constrain what anything can do with the infrastructure it authors — and Workato shipped an MCP registry to make that sellable to enterprise IT. With 523 improvements against 98 sparks, most of the week's volume was this kind of plumbing: attribution, audit, and the migration paths that let new capability land safely.

Leaders

Appwrite shipped four sparks in a near-daily infrastructure run, centered on rewriting its CLI from Node to Go as a single native binary — commands start in ~10ms instead of 200ms, with every flag and exit code deliberately unchanged so CI keeps working. Paired with a hosted remote MCP server and a Functions template that turns your own function into an MCP endpoint, it now sits on both sides of the protocol.

Rollbar reached the end of its AI arc: the Resolve agent entered public beta reading a production error's stack trace, telemetry, and root cause, then writing a fix, running the test suite in an ephemeral environment, and opening a pull request carrying the diff. It runs on every plan against a credit pool sold separately from tier — the packaging decision that made a free-plan rollout possible.

GitHub shipped Agent Plugins 1.0 with AWS, Anysphere, Microsoft, OpenAI, and Vercel as co-signers: a plugin built once runs in VS Code, Copilot CLI, the Copilot app, and any other compatible client GitHub does not own. Beneath the weekly model-catalog churn, portability is the actual bet.

HashiCorp turned its provenance thesis into shipped code. Packer v1.16.0 generates and verifies SLSA provenance for machine images natively, a week after HashiCorp argued that accountability — not authoring — is the control plane's product. It is the first place that argument becomes a capability a customer can use.

Sanity moved its MCP server faster than either Studio or the Media Library, and shipped Sanity New: sanity new provisions a project on temporary credentials that let an agent and a developer build for 72 hours before anyone signs up. The same window forced its JavaScript client and UI packages ESM-only onto Node 22.12 — capability and cleanup in one push.

Wildcards

Browser Use reoriented its commercial model around programmatic buyers: its API moved onto X402 micropayments with top-ups from one cent, and the paid boundary was redrawn so managing an open session is free and only opening a new browser costs money. Sub-cent top-ups and pay-per-browser-open only make sense if the thing buying is a program rather than a human with a credit card.

Auth0 extended its delegation model from API calls to full browser sessions — an authorized actor can now open a real web session as another user, carrying an RFC 8693 act claim that names who is actually driving. It is the same act-claim machinery Auth0 is building for agents, pointed here at human support staff, and it ships alongside a NIST-aligned password policy defaulted only for new tenants.

Themes that compounded

• MCP moved from integration to governed infrastructure: Appwrite, Sanity, and Workato all treated the server as a first-class, permission-checked surface rather than a bolt-on.
• Bounding the agent, not just enabling it: Zed sandboxed its agent's terminal commands and web fetches, while Rollbar and HashiCorp made agent-authored output auditable before it lands.
• Runtime AI, not only build-time: WeWeb let the apps it builds call OpenAI, Anthropic, and Gemini from backend workflows, and Agno let an agent query its own platform's traces in plain language.
• CI surfaces widening into automation runtimes: Bitbucket pushed Pipelines past CI/CD into agentic chores and merge queues.
• A parallel, unglamorous thread in scientific computing: scikit-bio and CP2K kept pushing GPU and array-API work through libraries that never assumed anything but NumPy on a CPU.

Watch this week

The near-term move is defaults following adoption. GitHub's opt-in OAuth token expiry and Auth0's new-tenant-only password policy both wait on migration data before they harden. Expect Appwrite to extend its rewrite to the runtime cold-start limits it keeps flagging, Rollbar's Resolve to reach past single-error fixes, and the MCP-governance pattern from Workato to extend attribution onto agent actions themselves. Wherever a product shipped agent capability this week, the next release is the accountability layer that makes it safe to leave running.