Salesforce Killed The Browser. Every Agent Runs Your CRM Now.
TL;DR
The real shift is from model wars to infrastructure wars — Nate argues the launches that matter now are the ones that change what your existing tools can reach, not the ones with the loudest benchmark chart or flashiest demo.
His five-question filter cuts through agent fatigue — evaluate every launch by asking whether it plugs into current tools, supports other agents, reaches important data, has an ecosystem, and lets you stack agents on top.
Salesforce Headless 360 is the sleeper release with the biggest implications — by exposing Salesforce capabilities as APIs, MCP tools, and CLI commands, plus 60+ new MCP tools and 30+ coding skills, Salesforce is effectively saying the browser is optional and every agent can become a CRM agent.
OpenAI Workspace Agents are best understood as reusable team work units, not universal replacements — they shine for recurring cross-tool workflows in ChatGPT or Slack like feedback routing, metrics reporting, and triage, but they lose when Microsoft 365 or Salesforce owns the core data graph.
Microsoft Copilot Wave 3 wins on native organizational context, not openness — Copilot co-work and Work IQ matter because they operate inside Outlook, Teams, SharePoint, Excel, PowerPoint, identity, and permissions, but that same closeness makes it weaker for external composability and engineering-heavy work.
Claude is increasingly a layer inside other products rather than just a destination app — Nate points to Claude Sonnet 4.5 in Salesforce, Anthropic tech inside Microsoft Copilot co-work, and Claude Opus 4.7 powering Perplexity Computer as evidence that the smarter question is no longer 'Should we switch?' but 'Which wrapper fits the job?'
The Breakdown
Agent launch fatigue is real, and the old comparison game is breaking down
Nate opens on the weekly flood of announcements — OpenAI workspace agents, Anthropic managed agents, Salesforce Headless 360, Perplexity Personal Computer, Moonshot’s Kimi K2.6 — and says the dominant reaction from operators isn’t excitement, it’s exhaustion. His core setup is that the useful question is no longer “what launched?” but “which of these deserves an afternoon of my team’s attention?”
The five-question filter: infrastructure beats flashy features
He lays out a practical filter: does it plug into your existing tools, can other agents build on it, does it access data you care about, is an ecosystem forming, and can you stack agents on top? The framing is sharp: “features commoditize, infrastructure compounds,” and migration into yet another destination app is usually the most expensive ask you can make of a team.
OpenAI Workspace Agents: from personal assistant to reusable team worker
Nate says Workspace Agents matter because they move agents from one-off chats into shared, scheduled, repeatable workflows that can run in ChatGPT or Slack. His examples are refreshingly boring in the best way — product feedback routing, weekly metrics reporting, risk screening, software request triage — exactly the stuff companies keep trying to automate but lose in the mess between docs, email, Slack, spreadsheets, tickets, and “somebody’s memory.”
Salesforce Headless 360: the browser stops being the center
This is the release he thinks people may overlook even though it could matter the most. Salesforce announced at TrailblazerDX that major platform capabilities will be exposed as APIs, MCP tools, or CLI commands, with 60+ new MCP tools, 30+ preconfigured coding skills, support for Claude Code, Cursor, Codex, and Windsurf, plus AgentExchange and a builder fund; as Parker Harris put it, “Why should you ever log into Salesforce again?”
The Claude-under-the-hood pattern shows up everywhere
A key detail in the Salesforce story is that Agentforce Vibes uses Claude Sonnet 4.5 by default, with GPT-5 available through multimodel support. Nate uses that to make a broader point: Anthropic increasingly looks less like a standalone destination and more like an embedded engine inside Microsoft, Salesforce, Perplexity, and now its own managed agent infrastructure.
Copilot Wave 3: unbeatable inside Microsoft’s graph, less compelling outside it
For Microsoft-native companies, Copilot co-work and Work IQ matter because they operate inside the full 365 context — email, meetings, chats, files, SharePoint pages, identity, permissions, and org structure. Nate’s distinction is useful: ChatGPT with connectors can access files, but Copilot is “operating inside the organizational graph,” which is a real moat even if the product is more closed and much less appealing for serious engineering teams.
Kimi K2.6 is impressive infrastructure, but mostly for teams building their own stack
Moonshot’s Kimi K2.6 gets attention because it’s an open-weights multimodal agentic model with a 300-subagent swarm and up to 4,000 steps, released under a modified MIT license. Nate’s point is that this is strategically important without being the default pick for most business teams: if you can self-host, inspect, fine-tune, and keep data in-house, it matters a lot; if you’re a typical Western business user asking what tool to adopt next week, it probably doesn’t.
Perplexity Personal Computer closes the local-work gap — and the real lesson is layering, not switching
Perplexity’s Mac rollout adds local file editing, local computer use, local browsing via Comet, voice orchestration, and deeper background control, with Claude Opus 4.7 as the default orchestrator. Nate says it’s best for research-heavy work that ends in an artifact — competitive intel, sales prospecting, market maps, document review, weekly ops reports — and closes by arguing that the market is moving toward layers: keep a default where it works, add specialists where they clearly win, and route work by task shape rather than chasing whichever agent had the loudest launch.