Day 97: Build vs Buy the Agent Runtime

Build vs buy the agent runtime is not ideological. Bedrock Agents, a custom loop on Step Functions, and IDE hooks (Cursor-style) optimize for different constraints.

⚡ TL;DR: Buy/managed when tool schemas are simple and AWS-native. Custom loop when you need strict dual control, twins, and audit. IDE hooks for developer UX — not for prod mutate paths.

Decision table

# ✅ Thin custom loop — still behind gateway
def agent_turn(state, model, tools, policy):
    plan = model.plan(state)
    for intent in plan.intents:
        policy.check(intent)           # Cedar/OPA
        result = tools.dry_run(intent) if intent.mutate else tools.run(intent)
        state = state.reduce(intent, result)
        if state.should_stop():
            break
    return state

❌ Rewriting a managed agent framework from scratch to avoid reading the docs — or adopting managed agents when you need byte-level audit of every tool.

Failure modes

Managed agents that cannot emit the audit shape compliance requires — you will rewrite later at 10× cost. Spike audit export on week one. Custom loops that reinvent tool JSON parsing bugs — reuse validated schemas either way.

Closing checklist

• [ ] Constraints listed before vendor demo
• [ ] Spike both paths on one real workflow
• [ ] Audit/dual-control requirements written
• [ ] Exit plan if managed lock-in hurts
• [ ] Gateway remains the front door either way

Series navigation

Day 96: Hiring and Team Topology for AI Platform · Day 98: Capstone Spec: Pick One Production Problem

Last updated September 11, 2026