My AI agent self-healing fix was embarrassingly simple once I stopped retrying the whole thing

Description and use

Solution: use a stable execution ID, persist state after each step, and retry only at the step level to avoid duplicate side effects and improve agent resilience.

Problem

Full-workflow retries caused duplicate work, increased costs, and fragile agents when a single step failed.

Outcome

The agent stopped repeating successful steps, avoided duplicate CRM updates, reduced token usage and latency, and became resilient to transient failures.

Workflow

• Assign a stable execution_id to each run
• Persist state after each meaningful step
• Retry only failure‑prone steps (e.g., HTTP calls)
• Use idempotency keys for side effects (CRM upsert, Slack post)
• Resume from the last completed checkpoint using the same execution_id

Evidence

The best ai agent self-healing pattern is not a smarter prompt. It is durable execution: save explicit state, retry only failure-prone steps, and resume with a stable execution ID.

Source

That one change killed most of the chaos.

Source