The CNCF Platforms White Paper defines platform engineering across dimensions like self-service, golden paths, documentation, and guardrails. Each dimension has a direct analog for AI agents.
| Dimension | Human Teams | AI Agents |
|---|---|---|
| Self-service goal | Reduce cognitive load, eliminate tickets | Reduce context window waste, eliminate ambiguity |
| Interface | Portal, CLI, UI, docs | MCP tools, APIs, structured schemas, tool descriptions |
| Golden paths | Opinionated templates, scaffolding, starter kits | Skills, agent prompts, workflow definitions, pre-built tool chains |
| Discovery | Service catalog, search, docs site | Tool registries, ToolSearch, capability advertisement |
| Documentation | How-tos, runbooks, READMEs | Structured context (CLAUDE.md), tool schemas, few-shot examples in descriptions |
| Cognitive load | Too many choices, context switching, tribal knowledge | Context window limits, token budgets, irrelevant context noise |
| Guardrails | RBAC, policies, compliance gates | Sandboxing, permission boundaries, tool allowlists, destructive action blocks |
| Toil | Manual deploys, ticket queues, config drift | Re-discovering context, re-reading files, redundant searches, hallucinated paths |
| Onboarding | Orientation docs, buddy system, ramp-up period | System prompts, CLAUDE.md, agent norms, skill loading |
| Collaboration | Slack, PRs, pair programming | Agent channels, MCP messaging, shared memory stores, handoff protocols |
| Scaling | Hire more people, cross-train | Spawn more agents, share skills, parallelize |
| Failure mode | Burnout, knowledge silos, bus factor | Context overflow, stale memory, tool misuse, cascading bad decisions |
| Cost model | Salaries, time, opportunity cost | Tokens, API calls, compute time, context window budget |