Why the retry loop is usually the expensive part of agent work

Short summary

When agents hit errors, the expensive part isn't the first failure—it's the retry loop that keeps running against the same problem without detecting it. Effective agent systems need bounded retries (max attempts, budget caps), clear verification criteria, and detailed receipts showing what was attempted and why it stopped. Stricter control systems stop earlier, cost less, and maintain operator trust better than agents that blindly push through friction.

• •Unbounded retries compound mistakes; retrying in an unchanged environment just repeats the same error
• •Control systems beat prompt engineering: set budget caps, max attempts, failure detection, and emit receipts
• •Early stopping with visibility costs less and preserves trust; real problem is failure classification, not smarter prompts

Generated with AI, which can make mistakes.