Make the retry safe and the result harmless
From Claude Certified Architect – Professional: Independent Exam Prep · 00:15:34
Show by Justin Thompson
On Rotterdam's second Monday live, the shift-change rush brings rate-limit errors the wrapper retries, then create-work-order calls that time out after the facilities system has already executed them, and the wrapper retries those too: thirty-seven duplicate work orders by eight o'clock. The same week a red-team day plants content at four entries (a retrieved procedure, a ticket field, a vendor MCP server's response and a third-party Skill), and two of them change what the assistant does. Read the trace for its signature and separate a wrapper retry from a model calling twice; match the control to the mechanism (honor retry-after, retry only what is safe to repeat, key every side-effecting call on its approval record, fall back without leaving the residency boundary); and inventory the entries where untrusted content reaches the model, keeping each one data with tagging, screening, allowlists, audits and version pins under least privilege. Original HarborWorks case aligned to Claude Certified Architect – Professional domains 3 and 5 (Production Reliability Controls; Prompt Injection Surface and Controls) in the December 7, 2026 guide, v1.1.Independent exam preparation, not affiliated with or endorsed by Anthropic; personal views, no employer endorsement. Original examples only, no actual exam items, no credential guarantee. This episode follows the Claude Certified Architect – Professional exam guide effective December 7, 2026 (version 1.1); check the guide that matches your exam date.Sources:Claude Certified Architect – Professional Exam Guide, v1.1 effective December 7, 2026 (Anthropic certification page)Rate limitsClaude API errorsHandle tool callsMitigate jailbreaks and prompt injectionsMCP connectorAgent SkillsHooks referenceSecurely deploying AI agents
More episodes of Claude Certified Architect – Professional: Independent Exam Prep
More episodes
- Pick the surface before the architecture
- Calibrate the judge before you move the model
- Put the control before the consequence
- Make obligations and unequal outcomes reviewable
- Leave a decision the next team can change
- Diagnose before declaring a winner
- Turn traces into an operating decision
- What must this service improve?
- Give each question the right evidence
- Enable the team and measure the workflow
PodSearch is a free, independent podcast search engine. Read about PodSearch, the privacy policy or contact us.