Weekly signal

This briefing covers the week 2026-11-02 through 2026-11-10 and highlights concrete platform releases and APIs that change how humans and agentic AIs coordinate work: Google’s ADK moved core agent plumbing from experimental to stable (including agent-to-agent protocols and safer shell execution), Cloudflare upgraded its Pi harness to 1.5 with new session/watch and max-turn controls, GitHub made Copilot available inside JetBrains Air via the Agent Client Protocol (ACP) so developers can run and review cloud-local agent tasks, and Hume launched an Audio Decision API that answers structured questions about audio (probabilities, not transcripts)—each change reduces friction or reshapes oversight points for human–AI workflows.

What changed

1. Google ADK v2.14.0 (released 2026-11-07) graduated A2A (agent-to-agent) integration, sandboxed execution environments, evaluation modules, and several tool integrations to stable — and added TeamsRunner and an OpenAI Decisions classifier for gating tool calls. The release also introduces breaking and behavioral changes (enum argument types, safer RunSkillScriptTool semantics to close a shell-injection vector, and session-format incompatibilities on rollback). This makes more of the ADK stack “production-ready” but requires careful upgrade planning.

2. Cloudflare Agents SDK ([email protected], published 2026-11-03) moved Pi harness to Pi 1.5, which batches persistence, replaces session.events with session.watch, adds session.continue and defaults.maxTurns, and establishes a compatibility boundary (sessions written by 1.5 can’t be read by older Pi). The change prioritizes throughput and bounded runs, but introduces a migration/rollback risk.

3. GitHub announced Copilot integration with JetBrains Air (2026-11-03): Copilot can be selected in Air’s agent picker to run tasks locally or in isolated cloud tasks billed to the user’s Copilot plan. Integration uses the Agent Client Protocol (ACP), exposing cloud-local execution, parallel sessions, and in-context review workflows inside IDEs. This tightens the loop between human review and agent execution.

4. Hume launched the Audio Decision API (2026-11-02): one-call, plain-language questions about audio (speaker count, laughter, frustration, etc.) that return thresholdable probabilities rather than transcripts or prose. It’s positioned for augmenting voice/agent decision points (e.g., “route this call to supervisor?”) and can replace multi-model audio pipelines with a single decision surface. Access is gated via waitlist / request.

What to do with it

1. Treat ADK v2.14.0 as a signal that core agent plumbing is stabilizing — but don’t upgrade blindly. Run a staged upgrade: test MultiRun evals, validate enum handling, back up sessions, and examine any flags you relied on because several “disable” switches were removed. If you rely on shell-execution tools, adopt the safer RunSkillScriptTool patterns now.

2. For Cloudflare-based agents, schedule a Pi 1.5 migration window: snapshot databases, test session.watch-based monitoring, set sane defaults.maxTurns for human attention budgeting, and update rollback playbooks because older Pi versions cannot read 1.5 sessions. Use the new continuation and max-turn controls to enforce human-in-the-loop checkpoints.

3. If developers use JetBrains IDEs, enable the Copilot-in-Air flow for code-review + run workflows; prescribe policies for cloud tasks, auditing, and who may run cloud agent sessions (since tasks bill against Copilot plans and run in isolated containers). Use ACP-compatible policy checks and require review gates for high-risk changes.

4. For voice/contact-center and multimodal agent integrations, prototype replacing multi-stage audio pipelines with Hume’s decision calls for routing/triage rules—but evaluate bias, legal/privacy implications, and how decision confidences map to human escalation thresholds. Because Hume returns probabilities (not transcripts), instrument explainability and fallback paths for humans.