Google Donates Agent Substrate and gVisor to the CNCF

Google is donating Agent Substrate and gVisor to the Cloud Native Computing Foundation. It also announced OpenRL, a project for orchestrating reinforcement-learning workloads. Agent Substrate separates an agent's state (actors) from pooled, pre-warmed compute (workers) on Kubernetes. Google notes that agents can sit idle 90–95% of the time, so tying each one to a dedicated pod wastes capacity. In Google's own figures, a suspended agent can resume in under 500 milliseconds.

gVisor adds a user-space kernel that intercepts system calls. That limits how much of the host kernel is exposed when an agent runs AI-generated shell commands. For stronger isolation, Agent Substrate also supports Cloud Hypervisor microVMs. Placing both projects under vendor-neutral governance signals that agent runtimes are becoming a separate infrastructure layer, distinct from serverless functions and long-running services.

Read more — InfoQ


Keep the Prefix Stable: Ollama and Together AI on Long-Horizon Agent Performance

An Ollama blog post from November 16, co-written with Together AI, argues that the prefix cache decides an agent's speed and cost with open models. Each agent turn resends the whole conversation, and any change early in the prompt invalidates the cache for everything after it. The practical rules:

  • Append new tool definitions rather than editing the initial tool list.
  • Keep system prompts static: no timestamps or git status.
  • Fix reasoning effort at the start of the session, because models such as GLM-5.3-Flash write it into the first system block.
  • Add explicit truncation notices for large tool outputs.
  • Give subagents either the inherited prefix or a short brief, so they reuse the cache.

On the serving side, the post describes routing requests by prompt_cache_key or prefix hash, offloading idle caches to CPU, and separating prefill and decode clusters. The same ideas appear in Hermes Agent v0.23.3, released this week, which added a prompt-cache lifecycle controller and cache-aware compaction. Prefix stability is becoming a design constraint for agent harnesses rather than a tuning detail.

Read more — Ollama


Harness Adapters: One LoRA per Agent Harness

Hugging Face's November 20 post on harness adapters trains one LoRA adapter per agent harness on Qwen3.8-27B, using TRL's AsyncGRPOTrainer. The harnesses are Pi 1.5.2, OpenCode 2.2.7 and Hermes Agent v0.22.6. The adapters were trained on 871 agentic coding tasks and evaluated on Terminal-Bench 2.1 and SWE-bench Pro. With thinking turned off, the Pi adapter raised SWE-bench Pro pass@1 from 40.0% to 64.0%. On tasks both models solved, it made 56–72% fewer tool calls.

The three adapters were then merged into a single "super-adapter" (LoRA rank 96, merged at weight ½). Even with thinking off, it beat the base model with thinking on in 5 of 6 harness-benchmark pairs. The finding is that an open model doesn't have to fit every harness equally well. A cheap adapter matched to a harness's tool schema and conventions can recover a large share of the gap to frontier models, and cut the number of tokens spent.

Read more — Hugging Face


Testing Agents Like Code: Copilot Evals and Datadog Online Experiments

GitHub Copilot Evals, in public preview since November 17, lets you write repeatable tests for skills, custom agents and MCP tools. An eval gives an agent a realistic task and checks both the result and the path the agent took to reach it. You can author evals in plain language or with the /eval command, approve each run, and compare runs before and after a change to catch regressions. The engine is Vally, an extensible local eval tool. Copilot Evals ships as a VS Code extension, a Copilot CLI plugin, and a marketplace entry in the Copilot app.

For production, Datadog added Online Experiments to its Agent Observability product. It splits live traffic between a control and a variant and compares LLM-judge quality scores alongside latency, tokens, cost and errors, with links to individual traces. In Datadog's example, switching from a GPT-5.6 to a GPT-6 variant cut estimated cost by 44.3% with nearly identical quality scores (3.84 vs 3.86). The traces, however, showed the new model passing invalid tool parameters more often, which the quality score did not catch. Offline evals and aggregate quality scores can miss failures in tool calls, so check the traces too.

Read more — GitHub Changelog


Vercel UI SDK: Headless Behaviors That Agents Can Drive Without Screenshots

Vercel released UI SDK in beta (0.1.0-beta.1, Apache 2.0). It is a TypeScript library of more than 50 headless behaviors, such as createPopup and createListbox, with adapters for React, Vue, Svelte, Solid and Web Components. Agents use createAgent to snapshot, query and act on components through a structured API, without parsing screenshots. registerTools can expose the same methods through WebMCP.

From a safety point of view, the useful part is that change callbacks receive a source of "agent". An app can therefore refuse or ask for confirmation on changes made by an agent, while letting the same change through when a human makes it. Vercel says the APIs may change before GA.

Read more — InfoQ


Netflix's Ontology-Backed Knowledge Graph as the Coordination Layer for Agents

Netflix described an ontology-driven knowledge graph of operational entities: applications, incidents, deployments, alerts and dependencies. It is built on RDF, OWL, SHACL and SPARQL and stored in QuipuDB, Netflix's RDF database built on redb. Enrichers populate the graph. Some are deterministic and some are LLM-based, and all are checked by strict SHACL validation. A correlation engine evaluates 13 signal types and produces ranked hypotheses about causes.

The graph powers AutoRCA, Netflix's internal agentic root-cause analysis system. Instead of passing loosely structured findings to each other, the agents share the ontology as a common, typed interface, and humans keep approval over any remediation. For teams building multi-agent systems, it shows a shared, validated world model replacing free-text messages between agents.

Read more — InfoQ


Safe & Secure AI Agent Practices

Google Agent Compute Engine: Credentials Never Enter the Sandbox

Google announced Agent Compute Engine, a managed service that runs AI agents in isolated microVMs and can checkpoint them so they can be suspended and resumed. The key design choice is that credentials, API keys and OAuth refresh tokens never enter the sandbox. An external Agent Gateway authenticates outgoing requests and injects the secrets. Google separates the "orchestration harness" from the execution sandbox and says this gives a deterministic audit trail that links each tool call from the harness to activity in the sandbox.

The service is not available yet. It has no published pricing, is targeted for preview "in the coming months", and runs only on Google Cloud. The pattern applies anywhere, though. If an agent's sandbox holds no usable secrets, a prompt-injected agent has nothing to steal.

Read more — InfoQ


Copilot Admins Can Block Standalone MCP Servers, Hooks, and Agents

GitHub added the managed setting strictPluginOnlyCustomization for Copilot on Enterprise Cloud. It requires that customizations come from plugins or managed settings rather than standalone user or project configuration. You can apply it to all five categories or pick them individually: skills, custom agents, hooks, MCP servers and executable Copilot extensions. For example, an organization can lock down MCP servers and executable extensions and leave skills, agents and hooks open.

The setting is not a marketplace allowlist. GitHub recommends pairing it with strictKnownMarketplaces to control which marketplaces plugins can come from. Together they close a common gap in agent governance: a repository-level MCP or hook config that runs on a developer's machine without review. Approved tools can still be distributed through a shared plugin library.

Read more — GitHub Changelog


Docker Sandboxes: Web Access Through a Proxy-Managed Key

Docker published a Firecrawl kit for Docker Sandboxes (docker.io/firecrawl/sbx-kit-firecrawl:1.0.0). It gives sandboxed agents web search, scraping and crawling while opening the sandbox's network allowlist to just one host, api.firecrawl.dev. Inside the sandbox, FIRECRAWL_API_KEY is a proxy-managed placeholder, and the proxy injects the real key on the way out. Under the balanced policy, requests to general websites still get a proxy 403. The agent can read documentation through Firecrawl without the sandbox being opened to the whole internet. The kit needs sbx CLI 0.46 or later and is in Early Access.

A companion post on Docker AI Governance points out a limit that matters. MCP tool policies, written in Cedar, apply only to traffic that goes through Docker's MCP Gateway. A direct MCP connection from inside a sandbox bypasses them. Teams that rely on gateway-level tool policy also need network policy that blocks direct MCP connections.

Read more — Docker Blog