Ten more things that go wrong when you build with a team of AI agents (part 2)

Description and use

The checklist and practices reduced server restarts, misattributed decisions, time zone errors, silent stalls, bottlenecks, rule conflicts, proposal overload, secret leaks, quota surprises, and cascading failures, improving reliability and predictability of the AI agent team.

Problem

AI agent teams often encounter subtle failures that waste time, misattribute decisions, and expose secrets, leading to costly surprises and bottlenecks.

Outcome

The checklist and practices reduced server restarts, misattributed decisions, time zone errors, silent stalls, bottlenecks, rule conflicts, proposal overload, secret leaks, quota surprises, and cascading failures, improving reliability and predictability of the AI agent team.

Workflow

• Assign each agent unique ports, accounts, and test data
• Log every decision with the actual author and timestamp
• Write dates in absolute form with time zone information
• Structure agent outputs so questions appear first and are tracked
• Create a coordinating agent for small, reversible decisions
• Validate new rules against existing ones before deployment
• Move non‑strategic proposals to a “not now” list
• Store secrets in local, ignored files and refuse to echo them
• Measure run costs, set budgets and alerts, and gate expensive actions
• Document all infrastructure connections, back up before changes, and test incrementally

Evidence

Agents compete for the same machine – a server one agent is using gets restarted by another.

Source

Agents put words in your mouth – a decision appears in a document with your name on it, but it came from another agent's suggestion.

Source

Time is harder than it looks – a test fails every night around midnight. A message says "tomorrow" and two agents read different days.

Source

Agents stall silently – an agent has been "working" for hours. In fact it asked a question at the end of a long output and is waiting for an answer nobody saw.

Source

The human becomes the bottleneck – twenty small questions a day: which word, which colour, retry or not.

Source

Secrets end up in the chat – an agent needs a key to continue, and the fastest way is to paste it into the conversation.

Source

Quotas and costs surprise you – in the middle of the month a limit hits: build minutes, an API quota, storage, a plan limit.

Source

One change ripples through everything connected – a repository moves, a key is rotated or a domain changes, and three other things break.

Source