5-Question Filter Cuts AI Agent Launch Noise
Evaluate agent launches with 5 questions prioritizing infrastructure: plugs into existing tools, buildable by others, owns key data, has ecosystem, stackable. Layer by task shape—don't switch providers.
Master the 5-Question Filter to Prioritize Infrastructure
Agent launches succeed when they enhance existing workflows, not demand migration. Apply this filter to any release:
- Plugs into tools your team already uses? Infrastructure extends agents into current environments like Salesforce or Microsoft 365; avoid new destinations that require data migration, as proven costly in SaaS history.
- Buildable by other agents? Open APIs, MCP tools, or SDKs (e.g., Claude Code, Cursor) make it infrastructure that compounds; closed products commoditize.
- Owns/accesses data you care about? Data access trumps model brilliance—a mediocre agent with full customer history outperforms one with empty context.
- Ecosystem forming? Watch marketplaces, SDKs, partner programs, and shipping cadence; one-off demos fade, ecosystems endure.
- Stackable with your agents? Composability multiplies value over adding isolated agents.
Launches passing all five deserve team time; others wait for Fridays. This shifts focus from benchmarks/demos to what expands reach and stackability.
Recent Launches: Winners Layer Data and Workflows
Salesforce Headless 360 excels as CRM infrastructure. Exposes all platform capabilities via 60+ new MCP tools, 30+ preconfigured coding skills, APIs, and CLI—agents access live org data without browser logins. Supports Claude Code, Cursor, Codex, Windsurf; Agent Exchange marketplace unifies ecosystem. Agent Force 5 defaults to Claude Sonnet 4.5. Passes filter fully: plugs into RevOps, open to external agents, owns CRM data, strong ecosystem (builder fund), fully stackable. Every agent now runs your CRM.
Microsoft Copilot Wave 3 dominates Microsoft 365 natives. Co-Work enables long-running multi-step agents (powered by Anthropic tech); Work IQ accesses full graph (email, meetings, chats, files, SharePoint, identity). Ideal for Excel/Outlook/Teams workflows with native permissions. Strong on data moat and governance but weaker on external composability/ecosystem openness—skip for cross-tool or heavy coding.
ChatGPT Workspace Agents fit shared recurring workflows. Cloud-based, Slack/ChatGPT-integrated, schedulable for team reuse (e.g., feedback routing, metrics reporting). Beats custom GPTs for cross-tool repetition but cedes to natives like Salesforce for CRM depth.
Perplexity Personal Computer suits research-to-artifact tasks. Mac app adds local file editing, browsing, voice; defaults to Claude Opus 4.7. Chains research/analysis/docs for intel/prospecting/reports. Moderate ecosystem; best for individual deliverables, not org-wide governance.
Kimi K 2.6 powers self-hosted dev teams. Open-weights (modified MIT), multimodal with 300-subagent swarms up to 4,000 steps; strong coding/agent benchmarks. For fine-tuning on own hardware, avoiding closed providers—not for hosted business teams lacking trust/governance.
Route Tasks by Shape, Layer Over Switching
Framing as 'switch to one agent?' misses the shift: agent market builds layers, not defaults. Route by task:
- Recurring cross-tool (Slack/email/docs): Workspace Agents.
- CRM/RevOps: Headless 360 (all agents now access).
- Microsoft-native: Copilot's graph.
- Self-hosted coding/swarm: Kimi.
- Research artifacts: Perplexity.
Claude embeds as engine (Salesforce Sonnet 4.5, Microsoft Co-Work, Perplexity Opus 4.7)—you're likely using it without direct switch. Wasted spend comes from forcing one tool everywhere; layer correctly to multiply agents across data/tools. Leaders stacking by task outpace launch-chasers.