What actually breaks when an AI agent runs unattended
Agents do not fail because the model is bad. They fail because state leaks between runs, retries are not idempotent, and nobody owns the alert. Here is the failure list we keep on hand.
ZetImagine Blog
Field notes on autonomous agents and Solidity
How we build autonomous AI agents and audited Solidity contracts — the trade-offs, the failure modes, and the parts that actually mattered in production.
-
-
Writing a Solidity contract: the checklist we actually use
A pre-deploy checklist that has caught real bugs on our own contracts — access control, token decimal traps, upgrade paths, and the arithmetic mistakes that survive a normal test suite.
-
Prompt caching is most of your bill, and most teams lose it by accident
A long conversation reuses a cached prefix only if the bytes before the new turn never change. One skill reload, one re-sorted memory block, one rebuilt system prompt — and you pay full price for the whole history again.
-
Designing agent tools the model actually calls correctly
Model error rates on tool calls are mostly interface design failures. Narrow tools beat clever ones, names carry the meaning, and every schema description should tell the model when not to use the tool.
-
Shipping agents to production without a compliance bill you did not plan for
The practical subset of AI governance that actually changes risk: what the agent can touch, what it writes down, who approves the irreversible parts, and how you prove it later.
-
Why your agent quietly gives up on the long task
Long work usually dies on a budget ceiling, not a capability wall. The three ceilings that end a real task — iterations, tokens, wall clock — and how to design around each one instead of hitting them.