Agent identity and MCP-native tooling went from individual bets to a pattern across the development stack this week.

The week in development

The overriding theme across development infrastructure this week was authorization: not just authentication, but the harder question of what a running AI agent is allowed to do, on whose behalf, and with what audit trail. Three separate products — Speakeasy, Workato, and Rivet — shipped different pieces of the same puzzle: agent-first credential management, per-tool access controls for MCP servers, and typed execution APIs that remove ambiguity from what an agent can invoke. The coordination is coincidental but the signal is directional.

The second pattern is search infrastructure getting AI-native by default. Manticore Search shipped chunked multi-vector embeddings as a first-class indexing strategy. These aren't bolt-on features — they're a rearchitecting of what a search engine does when the query is a vector and the document is a chunk of a larger object.

Leaders

Speakeasy gave AI agents their own first-class identity: dedicated API keys with required expiry, allow-only runtime policies narrower than the human developer's access, and explicit owners attached to each agent credential. This is the missing piece for enterprises trying to audit what their agents actually did. It ships alongside a tool that scans which AI devices are accessing which surfaces.

Manticore Search 29.9 ships chunked auto-embeddings with four chunking strategies (mean, fixed, recursive, sentence) and multi-vector storage via floatvectorarray. The practical upshot: native RAG pipelines without preprocessing infrastructure. This release removes a category of glue code that every team embedding long documents was writing themselves.

Workato shipped per-tool, per-group RBAC for MCP servers to GA — the first enterprise authorization layer for multi-tenant agent deployments. Teams can now specify which tools a given group of agents can call, enforced at the MCP server level rather than trusting downstream model behavior. This directly addresses the blast radius problem in agentic workflows.

Rivet added MCP exposure for its actor management surface and typed JavaScript/Python execution APIs for agents. The MCP addition means coding assistants can inspect and call Rivet actors directly. The typed execution APIs eliminate a category of unpredictability — agents running Node.js or Python through typed APIs instead of bash subprocess calls have far more deterministic behavior profiles.

WeWeb shipped direct OpenAI, Anthropic, and Gemini model calls from backend workflows alongside static pre-rendering for public pages. The model access is the leading signal: a no-code builder with first-party LLM calls in backend logic is now a competitor to custom-built AI app frameworks, not just internal tools.

Wildcards

Sanity deprecated its Studio context plugin and made the Context app the sole destination for AI context configuration. This is an architectural consolidation move — Sanity is betting that the AI context surface becomes the primary interaction layer, not the visual studio. The daily MCP server updates (32 releases in recent history) reinforce that this is an operational infrastructure bet, not a UI refresh.

Themes that compounded

• Agent identity infrastructure shipped from three angles this week: Speakeasy (dedicated credentials), Workato (per-tool RBAC), and Rivet (typed execution APIs with predictable blast radius)
• MCP-native access became the standard integration pattern: Rivet, Sanity, and Workato all shipped or deepened MCP surfaces this week
• AI-native search went from opt-in to default: Manticore Search ships chunked embeddings as a first-class indexing strategy, not a plugin
• WeWeb blurs the line between no-code builder and AI app framework with direct LLM calls in backend logic

Watch this week

Watch how Speakeasy's agent identity model gets adopted — if enterprises start requiring agent-specific credentials as a compliance posture, it changes the procurement conversation for every AI development tool. Also watch Workato's per-tool MCP RBAC: if this becomes a template for how enterprise platforms handle agent authorization, it gives Workato meaningful leverage in the automation space beyond its existing integration catalog.