•Exponential Partners

Latest in AI | 30 Oct

Three stories worth your time, then the rest.

1. Before adding another agent, test a skill.

Worth revisiting: Anthropic's 22 Oct guidance says one agent loading skills on demand often beat subagent designs in enterprise comparisons. Its 2 Sep commerce guide explains the handoff tax: lost context, more tokens and extra latency. Subagents still earn their place for measured context degradation, independent parallel work or difficult tool selection.

Takeaway: compare task quality, cost and latency before merging agents. Keep the specialists that earn their context.

2. Gemini 4 Carbon starts with restricted access.

Google's new model targets longer coding and cyber-defence tasks. Access starts with a subset of cyber defenders; wider business and developer access is still ahead.

Takeaway: test Google's benchmark claims on your own workload before planning a rollout.

3. Microsoft will run more of the agent machinery.

The Copilot harness is in public preview in Foundry Agent Service. Microsoft runs the loop and sandbox; customers configure tools and approvals. Hosted-agent long-running resilience is separately generally available.

Takeaway: buying a managed runtime still leaves you responsible for what the agent may do.

Also worth reading this week:

  • • Salesforce outlines agent evals and spending caps; check each release stage (27 Oct).
  • • OpenAI/ServiceNow's MCP plugin is available; deeper capabilities follow later (28 Oct).
  • • Oracle AI Workspace is in development to bring approved workflows into ChatGPT (27 Oct).
  • • Cohere's PitchBook connector requires an active licence and SSO (26 Oct).
  • • OpenAI plans to honour a shared AI-training preference signal in robots.txt in the coming months (30 Oct).

Sources:

On X:

Related Insights