Agent identity governance and dedicated agent memory both landed this week, while the agentic runtime layer gained BYOC, AI-generated app deployment, and cross-repo PR triage.

The week in development

Two foundational pieces of AI agent infrastructure shipped this week and deserve to be read together. Speakeasy gave agents their own identity — scoped API keys, upstream accounts, and audit log entries separate from human developers — and then immediately shipped the enforcement side: MCP gateway rules that refuse connections from agents on an organization's Shadow AI blocklist. Weaviate shipped Engram, a dedicated named agent memory product with its own configuration model. Neither is a feature addition to an existing human-facing product. Both are infrastructure that only exists because agents need it.

Below those structural moves, the week shows a sector in rapid surface expansion: embedded MCP servers in CMS and low-code platforms, a cross-repo PR triage queue that ranks every open pull request with machine-generated evidence, and a V8 isolate runtime for deploying AI-generated apps to end users. The development layer is acquiring agent-specific primitives faster than most teams are consuming them.

Leaders

Speakeasy shipped two releases that form a complete governance loop. Agent-as-principal gives each AI agent its own API key, scoped permissions, its own upstream account, and a discrete audit log entry — governance that was previously a log-search exercise becomes a first-class object. Shadow AI blocking then applies that identity: the MCP gateway reads the organization's Shadow AI blocklist in real time and refuses connections from agents on it. The identity primitive without enforcement is observability; the enforcement without identity is blunt. Speakeasy shipped both.

Weaviate released Engram, its first dedicated agent memory product: extraction topics, bounded scopes, and retrieval modes configurable per agent. This is a named product with a configuration model, not a memory pattern documented in a blog post. In the same week, HFresh shipped — a disk-based vector index that runs vector search with low heap overhead, changes the deployment economics for large collections, and maintains the index incrementally in the background. Both releases address the same constraint: agents at scale need memory infrastructure that doesn't require keeping everything in RAM.

Rivet shipped BYOC and Dynamic Apps in the same release cycle. BYOC lets enterprise customers run the Rivet control plane inside their own AWS or GCP VPC while Rivet handles management — data stays in the customer's cloud. Dynamic Apps introduces a V8 isolate runtime for deploying AI-generated applications directly for end users: apps generated in conversation, deployed with no infra, scaling to zero. These two moves serve opposite ends of the market — enterprise compliance on one side, low-friction AI app delivery on the other — and shipping them together signals Rivet is building a full-spectrum agentic backend.

CodeRabbit launched Triage this week: a cross-repository PR queue that ranks every open pull request by priority, with machine-generated evidence for each ranking. Engineering managers get a single view across all repos, with explanations for why each PR is where it is. The same release also shipped TypeScript configuration — replacing YAML with a type-checked format that adapts to PR context at runtime, supporting reusable fragments and conditional logic. Two distinct expansions — operational visibility and configuration expressiveness — landing together.

Sanity added workflow authoring to its MCP server: agents generate Sanity Workflow definitions from plain descriptions, validate them, and deploy new versions, with parallel tool calls improving throughput. SAML enforcement also shipped, letting enterprise admins require SAML-authenticated identities for all new members. The SAML move is table stakes for enterprise; the workflow authoring is the signal — Sanity is systematically wiring every CMS operation into the agent surface, not just content reads and writes.

Wildcards

Manticore Search had an unusually strong two-release week. Version 29.12 rewrites the bulk import path for real-time tables — large ingestion jobs now write directly to disk chunks and commit atomically, bypassing the RAM staging bottleneck that previously made high-volume ingestion expensive. Version 29.9 added multi-vector per-document storage and chunked auto-embedding with configurable chunking strategies. Together these releases move Manticore from a fast text search alternative into competitive territory for vector-heavy AI workloads — a positioning shift that wasn't obvious six weeks ago.

Themes that compounded

• Agent identity governance shipped at Speakeasy (scoped keys + Shadow AI blocking) this week — the first product to ship both the identity primitive and the enforcement layer in the same cycle.
• MCP servers extended into workflow authoring at Sanity and app building at Appsmith, taking the pattern from read-mostly to full-write in CMS and low-code platforms.
• GitHub Copilot added Claude Opus and memory-aware security autofix this week — the agent memory integration in particular mirrors what Weaviate is building at the infrastructure layer.
• Rivet's Dynamic Apps and Appsmith's MCP server both shipped runtime surfaces for AI-generated applications — one via V8 isolates, one via tool calls — in the same week.

Watch this week

Speakeasy's agent-as-principal model is the clearest articulation of agent identity governance shipped by any devtools vendor so far. Watch whether other API gateway and MCP infrastructure players respond with their own identity models, or adopt Speakeasy's framing as a reference pattern. On the memory side, Weaviate's Engram is in early availability; the configuration API and retrieval modes will matter more than the launch announcement — watch the documentation and community feedback over the next week to understand what the actual developer experience looks like.