AI made fixed-workflow SaaS obsolete. What replaces it is not smarter software. It is a harness: a seven-component system that makes a general-purpose language model behave like a reliable enterprise tool. Tomasz Tunguz breaks down exactly what that harness contains, and the specificity is the value here.
The seven components are context and memory, tools and action, orchestration and loop, state and persistence, sandbox and compute, observability and governance, and cost and workflow optimization. Each one is a distinct engineering discipline. Context is not one system. A radiologist's image retrieval stack is nothing like a paralegal's keyword search across a billion documents. Tools require a registry, argument validation, approval gates, and clean failure handling. MCP is named as the connective tissue. State means a crashed agent resumes at step 8, not step 1. Observability means evals catch regressions before customers do. The seventh component, cost and workflow optimization, is about architectural judgment: which model tier fits each step, and what belongs in memory versus hardcoded skill.
The competitive implication is direct. The major labs will dominate the markets they prioritize. That concedes thousands of vertical markets to startups. When every company runs the same model, the harness is the product. Read the original for the full breakdown of each layer, because the details inside each component are where the actual startup map lives.
[READ ORIGINAL →]