Lessons from building an enterprise agentic system

The mindset shift

• Traditional SaaS organizes around CRUD & data. Agentic software organizes around decisions. Data still exists, but it's a byproduct of work getting done, not the center of gravity.
• Every line of code is a constraint on the AI. Hard-coded workflows cap what the system can figure out on its own. The less you predetermine, the more effective the agent becomes.
• Design around outcomes, not departments. "What does success look like?" is a better starting question than "what does the workflow look like?"

The core tension

• The determinism paradox is the central tension. You need predictability, but restriction kills performance. The answer is constrained autonomy: control the sandbox, not the steps inside it.
• Sub-agents exist for bounded goals, not features. "Credential this nurse" is a sub-agent. "Credentialing module" is traditional thinking.

The architecture

• One master agent per role, not per function. A recruiter and a coordinator see different tools, different sub-agents. The role defines the permission boundary.
• Everything is a tool with an input/output contract. Platform capabilities, integrations, actions. Tool allocation to agents is the permission model.
• Agents need infrastructure to stay healthy. Vector search, context compacting, self-wake, per-task knowledge bases. Without this, long-running jobs fall apart.

The unlock

• Evals are agents that QA other agents. Standard input, expected output. This is how you get deterministic behavior without hard-coding the logic.
• Tool-shaped architecture compounds. New integrations become tool definitions, not projects. Every agent benefits immediately. This was the unexpected unlock.