Most important take away

The agent market is no longer a switching question (Claude vs. ChatGPT vs. Copilot) but a layering question — the launches that matter are infrastructure plays that expose data and let other agents stack on top, not standalone destinations. Use a five-question filter on every launch (existing tools, openness, data access, ecosystem, stackability) and route each kind of work to the tool whose data graph and workflow surface fits it best.

Summary

Actionable insights for leaders and operators drowning in agent launches:

• Adopt the five-question filter for every agent launch before spending team attention on it:
  1. Does it plug into the tools my team already uses, or demand migration?
  2. Can other agents build on top of it, or is it closed?
  3. Does it own or access data I care about?
  4. Is an ecosystem (marketplace, SDKs, partners, cadence) forming around it?
  5. Can I stack my own agents on top?
• Stop asking “should I switch?” Start asking “which wrapper around my model fits this job?” Anthropic’s Claude now appears in three forms: direct (Claude/Claude Code), embedded inside other vendors (Salesforce agentforce vibes uses Sonnet 4.5 default; Perplexity Personal Computer uses Opus 4.7 default; Microsoft Copilot Co-work uses Anthropic tech), and managed agent infrastructure. The same logic applies to ChatGPT, Copilot, and Gemini.
• Route work by shape, not by default tool:
  • Recurring, shared, cross-tool team workflows in Slack/ChatGPT → ChatGPT Workspace Agents.
  • CRM/revops work where data lives in Salesforce → use Headless 360 so the agents you already have can act inside Salesforce (no need to adopt agentforce as a destination).
  • Microsoft 365-native work (Outlook, Teams, SharePoint, Excel) → Copilot Co-work + Work IQ, because the data graph wins.
  • Frontier coding and novel reasoning → stay in Claude Code, Codex, Cursor — model-centric, not wrapper-centric.
  • Self-hostable open-weight agent infrastructure for dev teams that need data sovereignty → Kimi K2.6 (skip the hosted Kimi product for sensitive business work).
  • Research-heavy deliverables (competitive intel, market maps, prospecting, financial analysis) → Perplexity Personal Computer, now with local file access on Mac.
• Don’t try to make one license cover every job class — the real cost is mis-routed work, not buying a second specialist tool.
• Treat Salesforce Headless 360 as the sleeper announcement: 60+ MCP tools, 30+ pre-configured coding skills, support for Claude Code/Cursor/Codex/Windsurf, an experience layer that decouples agent logic from output surface, and Agent Exchange marketplace. If your revops runs on Salesforce, every agent you already use just got CRM access.
• Recognize Anthropic’s enterprise pattern: be the model layer inside other people’s products rather than a single destination. This means you may already be a Claude user without “switching.”
• Career advice signal: “The judgment to route work based on the shape of the task is the new literacy of the agent era.” The differentiating skill for operators and leaders going forward is not picking one default agent — it’s building taste for matching job shape to tool layer (model vs. wrapper vs. data graph vs. open infrastructure). Teams that learn to layer deliberately will compound; teams chasing the loudest launch will fall behind.
• Switching costs are real: prompts, memory, and team habits don’t port cleanly. Keep your default; add specialists only where the win is clear.
• Heuristic for the rest of the year: filter for infrastructure over features, ecosystems over demos, stackability over walled gardens, data access over benchmark charts.

Chapter Summaries

The exhaustion problem. Every week brings another agent launch (OpenAI Workspace Agents, Anthropic managed agents, Salesforce Headless 360, Perplexity Personal Computer on Mac, Kimi K2.6 with 300-agent swarm). Leaders are exhausted, not excited. The real question is which deserves an afternoon of attention. The conversation has quietly shifted from model quality to infrastructure.

The five-question filter. (1) Plugs into existing tools or demands migration? (2) Open to other agents or closed product? (3) Owns/accesses data you care about? (4) Ecosystem forming (marketplace, SDKs, cadence)? (5) Can you stack your agents on top? Most launches fail; the survivors are worth real evaluation.

ChatGPT Workspace Agents. Shared, codex-powered team agents that run in cloud, work across connected tools, surface in ChatGPT or Slack, can be scheduled, and are built for repeatable workflows. Shifts the model from “personal assistant” to “reusable team work unit.” Passes the filter for recurring cross-tool team workflows; not the answer when the work is deeply native to Salesforce, Microsoft 365, or frontier coding.

Salesforce Headless 360. The sleeper announcement. Every Salesforce capability exposed as API, MCP tool, or CLI command. 60+ new MCP tools, 30+ pre-configured coding skills, support for Claude Code/Cursor/Codex/Windsurf, experience layer that separates agent logic from output surface, Agent Exchange marketplace, Builders Fund. Salesforce isn’t shipping an agent — it’s becoming infrastructure under the agent economy. Scores extremely high on the filter. Sleeper detail: agentforce vibes uses Claude Sonnet 4.5 as default, GPT-5 optional — part of Anthropic’s pattern of being the model layer inside other vendors’ products.

Microsoft Copilot Wave 3. Co-work brings long-running multi-step agent execution to Microsoft 365, built with Anthropic. Work IQ provides the data layer (email, meetings, chats, files, SharePoint, identity, permissions, org context). Strong on Microsoft 365 data access, native permissions, enterprise governance. Weak on openness to external agents and ecosystem energy. Right for Microsoft-native non-engineering work; wrong for cross-ecosystem or coding-heavy teams.

Kimi K2.6 (Moonshot). Open-weights under modified MIT license. Native multimodal agentic model with 300 sub-agent swarm coordinating up to 4,000 steps. Strong coding/agentic benchmarks. Matters for dev teams building self-hosted agent infrastructure with data sovereignty needs. Does NOT pass the enterprise filter — no work graph ownership, no native 365/Salesforce integration, no Western enterprise connector story. The hosted Kimi product is not the answer for sensitive business work; the open weights are the real story.

Perplexity Personal Computer (Mac). Adds local file editing, local computer use, local browsing, voice orchestration, background work. Claude Opus 4.7 is the default orchestrator. Moves Perplexity from “chatbot that searches” to “digital worker that researches, reasons, browses, edits, creates artifacts.” Passes for research-heavy deliverable work (competitive intel, market research, prospecting, financial analysis, doc review). Fails for shared recurring team processes or work deeply native to Microsoft 365 or Salesforce.

The routing pattern. Don’t force one product to do every job — assign work to the tool whose graph and workflow surface fit the task shape. The expensive thing isn’t owning multiple tools; it’s routing work incorrectly.

The switching question is wrong. The market isn’t moving toward one default agent — it’s moving toward layers. You may already be a Claude user via Copilot, Perplexity, or Salesforce without “switching.” Three questions instead: (1) When to stay in your default direct product? When the model is the center and the wrapper adds little. (2) When to use a different product running the same model? When that wrapper gives data access or workflow integration you can’t reproduce. (3) When to use a different model entirely? When the surrounding product matters more than marginal model quality.

Switching costs and the new literacy. Prompts, memory, and team habits don’t port cleanly. Don’t switch casually — layer deliberately. The judgment to route work by task shape is the new literacy of the agent era. Filter for infrastructure over features, ecosystems over demos, stackability over walled gardens, data access over benchmarks. Teams that learn to route across layers will compound faster than teams chasing the loudest launch.