Agent Collaboration Weekly AI News
November 9 - November 17, 2026Weekly signal
Between Nov 9–17, 2026, agent collaboration moved from conceptual demos and point solutions toward platform features, cross‑industry standards, and concrete security guidance. Google published production‑scale multi‑agent results and a multimodal benchmark; OpenAI documented multi‑agent primitives and signalled an operational pivot by deprecating its visual Agent Builder; the community pushed Agent Skills + MCP into vendor‑neutral governance; Adobe announced Creative Agent orchestration across Creative Cloud; and Microsoft released research showing a new weight‑space backdoor for agents. These items together change how you should design, test, govern, and secure cooperating agents.
What changed
Antigravity /teamwork: Google published an in‑depth engineering post describing Antigravity’s /teamwork multi‑agent harness and real problem wins across mathematics, chemistry, systems software and ML engineering. /teamwork is explicit about engineering patterns: partitioning context so each specialist agent receives only the files and tools it needs; coordinators that spawn and synthesize subagent outputs; critics and falsifiers that run formal checkers, test suites or hardware benchmarks; and fresh successor agents to avoid accumulation of spurious internal state. Google reports Lean‑verified proofs and production infrastructure wins (e.g., memory savings, Borg scheduling improvements) — demonstrating that multi‑agent orchestration can produce verifiable, auditable outcomes at scale. For builders this is a practical design reference for reliable team‑style agents.
Multimodal Knowledge Agent (MKA): Google DeepMind and Artificial Analysis launched MKA, a 200‑task benchmark purposely built around the multimodal, multi‑file projects knowledge workers actually do (decks, blueprints, spreadsheets). MKA runs inside an open harness (Stirrup) and uses an agentic autorater that inspects files and applies an expert rubric. The benchmark highlights that agents’ performance differs dramatically in turn efficiency and visual reasoning, and it forces evaluation of end‑to‑end workflow quality rather than single prompt outputs. If you’re building collaborative agents that must handle documents, images and spreadsheets jointly, adopt MKA‑style tests to avoid surprises in the field.
OpenAI: platform multi‑agent primitives + Agent Builder deprecation. OpenAI’s Responses and Agents docs now expose hosted multi‑agent orchestration actions (spawn_agent, send_message, wait_agent, interrupt_agent, followup_task, list_agents) and examples for delegating parallel reviewers and reconcilers. The docs make it practical for a root agent to spin up focused subagents and synthesize results, and provide guidance on when parallel delegation helps (independent, bounded subtasks) and when it hurts (shared mutable state, small linear chains). At the same time, OpenAI is deprecating the Agent Builder visual canvas with a scheduled shutdown on Nov 30, 2026; teams using the visual flow should migrate to the Agents SDK / ChatKit or export code. Operationally this accelerates programmatic, code-first agent orchestration while raising an immediate migration task for some developers. Pay attention to OpenAI’s developer‑message controls (you can require explicit consent before spawning subagents).
Standards: Agent Skills + MCP -> AAIF. The ecosystem took a major interoperability step when Agent Skills (procedural skill packaging) entered the Linux Foundation’s Agentic AI Foundation alongside MCP (Model Context Protocol). With core maintainers from Google, Anthropic, OpenAI and others, Agent Skills + MCP under neutral governance reduce vendor lock‑in and make a deployable pattern for ‘skills + tool APIs’ that agents can discover and load progressively. For multi‑agent orchestration this solves the “N‑by‑M” tool problem: skill manifests let agents fetch procedural instructions only when needed, while MCP provides a common runtime contract for calling live tools. This is the single most important move for long‑term cross‑harness collaboration.
Adobe: agentic creative workflows. At Adobe MAX Adobe announced the Adobe Creative Agent and local MCP integrations for Creative Cloud apps, letting agents orchestrate Photoshop, Illustrator, Premiere, After Effects and more. Adobe’s approach shows industry appetite for application‑level MCP integrations: agents will not only suggest outputs but can execute multi‑step editing pipelines across file formats and pro tools. For creative teams, this is a clear example of agents collaborating across specialized tools and manifests how MCP opens “hosted tool” surfaces to third‑party agents.
Security: Microsoft’s Lazarus / sleeper‑agent disclosure. Microsoft Research published threat findings showing a technique to turn a capable agent model into a sleeper agent by surgically editing a handful of dormant neurons (no retraining required), making it behave normally in evaluations and only trigger malicious payload behavior when a trigger appears. Microsoft couples the disclosure with detection and mitigation guidance (weight‑level audits, provenance, red‑teaming, permissioning and monitoring). For any multi‑agent system that runs third‑party or open‑weight models, this materially raises the bar for model provenance and runtime hardening: model weights must be treated as first‑class auditable artifacts and agent permissions must be minimized.
Implications and how these pieces fit together
-
Design patterns are converging. Antigravity’s partitioning + coordinator + critic pattern is now echoed by platform primitives (OpenAI) and standard packaging (Agent Skills). That means the community is converging on the same architecture: small, focused agents; explicit orchestration primitives; verified critic/audit steps; and incremental skill loading for context efficiency. Treat this as the default architecture for complex automation.
-
Evaluation must become workflow‑level and multimodal. Benchmarks like MKA flag that single‑turn evaluations miss real failures: agents must be tested across files, tools, and multi‑turn budgets. Add end‑to‑end, rubric‑based evals to CI for any agent that collaborates across documents or tools.
-
Interop reduces integration cost. With MCP + Agent Skills in an industry foundation, expect more off‑the‑shelf skills, MCP tool servers, and cross‑vendor agent plugins — lowering the plumbing cost of building cooperating agents. Plan to adopt or bridge to MCP/Agent Skills for long‑term portability.
-
Security changes your supply‑chain model. Microsoft’s research means you can’t assume a model that passes functional evals is weight‑clean. Add provenance checks, artifact signing, and tooling to detect weight‑space edits; limit agent permissions (principle of least‑authority) and require human checkpoints for consequential actions.
-
Operational timing: OpenAI’s Agent Builder shutdown (Nov 30, 2026) is imminent. Teams still using the visual canvas should export workflows and plan migration, and teams using OpenAI multi‑agent should codify developer messages that restrict when subagents spawn.
What to do with it (practical next steps)
- Architecture and small experiments
- Prototype a coordinator + 3‑specialist subagents pattern for one internal use case (code review, incident triage, or multi‑file report). Enforce tight input contracts (only the files each subagent needs). Run parallel vs. sequential versions and measure wall‑clock and token cost. Use the multi‑agent primitives if your platform supports them.
- Evaluation & CI
- Add at least two MKA‑style tests: (a) a multi‑page PDF + spreadsheet synthesis task, and (b) an image+spec engineering task. Measure turn counts, token use, and autorater agreement to detect regressions. Don’t rely solely on single‑turn NLU benchmarks.
- Interop & tooling
- Adopt Agent Skills (SKILL.md) for repeatable procedural knowledge and publish internal skill registries. If you expose tools, implement an MCP server or bridge to MCP for discovery and consistent schemas. Lockdown how skills are discovered and versioned.
- Security & governance
- Treat model weights as artifacts: require provenance, signatures, and an auditable acceptance process before a model is allowed to run with high‑risk permissions. Run weight‑level and behavioral red‑teams on models (including fuzzing for trigger phrases). Limit agent privileges (no owner credentials by default) and require human approvals for high‑impact tool calls. Follow Microsoft’s mitigation checklist where applicable.
- Operational housekeeping (if you use OpenAI)
- Export Agent Builder workflows now and schedule migration to Agents SDK / ChatKit before Nov 30, 2026. Add developer messages that forbid proactive subagent spawning unless explicitly requested. Update runbooks to account for increased token usage with parallel subagents.
- Product strategy
- If you build platform integrations, prioritize MCP-compatible connectors and a skill‑publishing UX: it’s where cross‑vendor demand is going (Adobe, Google, MCP). Offer sandboxed execution and audit logs for every tool call to make your platform attractive to cautious enterprises.
Quick source map
Key primary reads this week: Antigravity /teamwork (deep engineering patterns and results); the MKA benchmark and technical report (multimodal workflow evals); OpenAI multi‑agent docs and the Agent Builder guide/deprecation note (primitives + migration); Google Open Source blog on MCP + Agent Skills joining the AAIF (standards & governance); Adobe’s Firefly / Creative Agent announcements (app integrations + MCP); Microsoft Security Blog on sleeper agents / Lazarus (weight‑space backdoor disclosure and mitigation).
Stop reading agent demos. Give one a job you repeat every week.
Describe the work, test the first result, and keep the agent available without running your own server.
Plans from $130/month. Cancel anytime.
Hosted agent
OpenClaw or Hermes