Agentic AI and MCP for Product Managers
Loops, Retries and Runaways
The specific failure that makes agents expensive rather than merely variable is the loop that doesn't terminate. These have recognisable shapes.
Pattern | What It Looks Like | The Fix |
|---|---|---|
Repeat loop | Calls the same tool with the same arguments over and over | Detect repeats; stop after two identical calls |
Ping-pong | Alternates between two tools without progressing | Track whether new information is arriving |
Retry storm | A tool fails; the agent retries indefinitely | Cap retries per tool at two or three |
Scope creep | Starts solving a related problem it wasn't asked about | Tighter goal statement; step cap |
Polite stall | Keeps saying it will check one more thing | Hard step cap. This is the one caps exist for. |
Caps Are Not a Safety Net, They Are a Feature
Teams treat step and cost caps as emergency brakes. Treat them instead as a normal part of the design, and design what happens when one trips. A well-designed cap produces a useful handover, 'I got this far, here's what I found, here's what I couldn't determine'. A badly designed one produces an error message and discards everything the agent learned.