LoginSign Up
← Back to DevOps
Weekly · DevOps · Week of August 3, 2026

Agents get their own identity, runtime, and audit trail as GitHub retires Models and ships stacked PRs

agent-identitymcpagent-governanceagent-runtimedeveloper-platformssafer-defaults
Y
Generated 1h agoDrawn from 13 products

The week in development

The loudest move this week came from GitHub, and it pointed two directions at once. GitHub Models is fully retired as of July 30 — playground, model catalog, inference API, and bring-your-own-key all gone for every customer, with no successor named — while stacked pull requests entered public preview, turning the unit of review from one large diff into an ordered series of dependent layers. Read together, the two say the same thing: GitHub is conceding raw model access to the model vendors and reinvesting in the surfaces it owns outright — the repository, the review flow, and the enterprise controls that gate everything else. Copilot code review reaching GA with agent skills and MCP, plus npm publish-time malware scanning, round out a week that reads as narrowing where GitHub competes rather than widening it.

Underneath that headline, the sector spent the week wiring identity, runtimes, and governance around software that now acts on its own. Auth0 shipped agent identity primitives before most of the category admits it needs them; Speakeasy and Workato built the control and build planes for the coding agents teams already run; and Rivet turned its agent sandbox into a programmable runtime other platforms depend on. The common thread is that “agent” stopped being a demo and started being an operational surface — something you authenticate, meter, isolate, and audit like any other production system.

Leaders

GitHub carried the highest signal in the window, with five sparks against nineteen improvements. Beyond retiring Models and previewing stacked pull requests, it moved Copilot code review's agent skills and MCP support to GA across Pro through Enterprise, so reviews can pull in repository conventions and external systems instead of reasoning from the diff alone. The Copilot line has settled into a reach-then-govern rhythm: ship broadly, then follow with policy targeting, defaults, and device scoping.

Auth0 is staking out identity for agents while the primitives are still in Early Access. Agents as Principal gives each agent its own identifier, credentials, and audit trail instead of a borrowed client or user account, and Token Vault Privileged Worker lets a background agent pull a user's third-party tokens — Gmail, Drive, Slack — with no active session, by registering its own credential. Around that core, the B2B line kept filling gaps: org-scoped roles, Enterprise Connect federation, and organization-to-application entitlement checks.

Speakeasy is building the control plane for the coding agents companies already deploy — Claude Code, Codex, Cursor. Its hooks now detect when an agent activates a skill and upload the manifest, so the skills registry fills itself from live sessions with real summaries and versions rather than name-only stubs. Alongside it, per-person spend budgets can now block rather than only warn, and blocked MCP calls explain their denials.

Workato made AIRO generally available worldwide: a multi-agent system that plans, builds, and debugs automations from a plain description, reachable in-product or over MCP from Claude and Cursor. It reads job logs to diagnose failures and fix them in the same conversation, and respects RBAC with no extra configuration. Packaged vertical agents — IT Support Genie, EDI Genie — sit on top as the commercial wedge.

Rivet consolidated around agentOS, its virtual operating system for agents built on WebAssembly and V8 isolates. A new execution API runs JavaScript and Python directly instead of shelling out to bash, and a package registry lets an agent's tools be installed by name. Notably, Vercel's Eve and Flue both shipped agentOS support within two days of each other — the runtime is becoming a layer other vendors build on, not just a Rivet feature.

Wildcards

Weaviate is repositioning a vector database as agent memory. Engram, its managed memory and context service for agentic applications, reached general availability — packaging memory as a service instead of leaving developers to assemble it from collections, schemas, and retrieval code — while release 1.38 promoted the HFresh disk-based index and a built-in MCP server to GA and rebuilt async replication underneath.

Rollbar, at the low end of the velocity range, made a pricing move worth watching: AI Root Cause Analysis follows session and propagation IDs across services to trace a downstream symptom to its upstream cause, and it now reaches the free tier through a $5/month credit subscription rather than a plan upgrade. Metering AI debugging by credit is a different commercial shape than bundling it into a seat.

Themes that compounded

  • Agent identity as its own type. Auth0 gave agents distinct principals and credentials, and Kinde shipped Secure MCP so customers can hand their own users an authenticated MCP server — authentication is moving from the human to the agent.
  • MCP as a first-class product surface. Appwrite moved its MCP server to a hosted mcp.appwrite.io with browser OAuth, Workato exposed AIRO over MCP, and Weaviate built one into the database itself.
  • Governance following capability. Speakeasy added blocking spend budgets and denial reasons; Agno shipped v2.7.0 with one auth layer across every transport, scope-based authorization, and a CI-gating eval suite — agent frameworks growing the properties of ordinary server software.
  • Build layers going swappable. Nuxt 4.5 added Rspack 2 alongside Vite 8 and shipped its own docs-grounded agent, treating the bundler as a dependency rather than a bet.
  • Safer defaults over new features. Supabase stopped auto-exposing new public-schema tables to its Data and GraphQL APIs, and PlanetScale added enforced resource budgets on Postgres query traffic.

Watch this week

Watch whether the agent-identity and agent-runtime work moves from Early Access toward general availability and gains scoping controls — Auth0 needs policy layered on agent identity, Appwrite needs per-resource OAuth scopes on its hosted MCP endpoint, and Rivet has state and data access as the next gap in the agentOS stack. On the platform side, GitHub's stacked pull requests are the item to track for API and CLI support ahead of GA, since that is where the review-flow bet actually lands for developers.