Weekly signal

This week (covering Oct 26–Nov 3, 2026) centered on turning multi‑agent research and standards into production orchestration and governed harnesses. Major enterprise vendors shipped orchestration layers that let humans delegate outcomes (not just prompts), an open‑source harness appeared to standardize agent plumbing and safety checks, and new multi‑agent research continues to shrink the communication/compute cost of collaborative pipelines. These moves make agent-to-agent coordination a practical engineering concern — not a purely academic one — and shift attention to governance, observable protocols (A2A/MCP), and harness-level testing before deployment.

What changed

1. Oracle announced Fusion Companion and Fusion Agentic Home (Oct 26, 2026): an orchestration layer that interprets user objectives, picks and coordinates workflows, approvals, and AI agents across Fusion Applications, and lets external assistants (e.g., ChatGPT) trigger ERP/CRM/HR tasks while enforcing enterprise permissions and audit controls. This is positioned as a first‑class way to run multi‑agent, outcome‑driven work in enterprise systems.

2. MetaHarness (open source) launched at GitHub Universe (Oct 28, 2026): Cognitum One released MetaHarness — a factory for agent harnesses that scaffolds command‑line tooling, scoped memories, witness‑signed releases, red/blue testing, and an automated "darwin" test/promotion loop for harness variants. It’s explicitly about giving teams a repeatable, auditable harness for agents.

3. New research on relay/relay‑style multi‑agent collaboration (Oct 26, 2026): the arXiv paper “Multi‑Agent Relay Thinking” shows relay collaboration patterns that combine heterogeneous agent strengths while reducing token/coordination overhead vs. standard multi‑agent ensembles — important for cost and latency budgets in agent teams.

4. Platform capabilities continue to adopt multi‑agent features and A2A/MCP interoperability: platform changelogs and documentation show broader multi‑agent support (OpenAI multi‑agent tooling and model updates), and the Agentic AI Foundation / industry projects are converging on registry and protocol building blocks (A2A, MCP) for agent-to-agent calls and governance. These are the plumbing that makes safe cross‑vendor collaboration realistic.

What to do with it

1. If you run enterprise workflows: map outcome paths, not prompts. Design an orchestration layer or adopt Fusion Companion‑style controls — explicit permission checks, identity anchors for agents, and audit trails before you delegate writes to ERP/CRM systems. Start with low‑risk test cases (status updates, ticket routing) and instrument governance.

2. If you build agents: adopt a harness pattern. Use or evaluate MetaHarness‑style scaffolding (scoped memory, signed releases, automated red/blue tests) to make multi‑agent coordination repeatable and auditable. Bake in promotion gates and deterministic replay for post‑mortem.

3. If you research or optimize costs: try relay/heterogeneous pipelines (relay thinking) for tasks that require specialist subagents — you can often cut token overhead and latency while keeping quality competitive. Benchmark mixed-model chains (specialist → integrator) against flat ensembles.

4. For platform architects and standards teams: prioritize A2A/MCP compatibility, observability endpoints, and auth patterns now. If agents must call each other across vendors, you need standard consent, capability advertisement, and revocation flows in place before scaling.