We’ve been testing AI agents in production since around April 2025, and what’s already working well is the shift away from rigid trigger action workflows toward agent-led orchestration.
Instead of chaining dozens of conditional steps, we’re using agents to interpret context, decide the next action, and only fall back to deterministic logic for guardrails (validation, approvals, cost limits, and failure handling).
In practice, this has reduced workflow complexity and made automations far more resilient to edge cases and changing inputs.
For those deploying agents today: how are you structuring the balance between agent autonomy and manual or rule-based oversight? What decisions do you still refuse to fully delegate?