Skip to content

Operating agents and proving a recovery

Answer these formative prompts in your own work record before you open the public answer guidance. The numbers are fresh course-made instances, not the lecture's records. Each answer should name the line, field or run that another reader could check.

  1. An agent's report says, "Every call I made was allowed by a grant." One of its grants lets code.read reach every repository, and the task needs one. Is the report enough to accept the session? Which comparison is missing?
  2. A gateway policy allows 2000 tokens per run. Each request reserves 900 and each reply uses 600. Which request is refused, and how many tokens are unspent at that moment? Why is it refused although tokens remain?
  3. A run has a cap of 12 units. Each request reserves 5 and each reply costs 2. How many requests are sent before one is refused? Check your answer with python lectures/week-11/fixture/budget_ledger.py --cap 12 --reserve 5 --actual 2 from the course checkout.
  4. A request's reply is lost. Why does the budget keep the whole reservation, and what would go wrong if it released it?
  5. Three changes were worked on. Change y1 was accepted after runs of 4 units (checks failed) and 4 units, with 12 review minutes. Change y2 was rejected after one run of 2 units and 8 review minutes. Change y3 was accepted after runs of 6 units and 2 units (timed out), with 20 review minutes. A summary says, "A change costs about 4 units." What is the cost per accepted change in units and in review minutes?
  6. A rule route passes 47 of 50 frozen cases and a model route passes 46. The rule fails one case that the model passes. Is the rule route good enough to replace the model from these rates alone? What is the next check?
  7. A draft root-cause analysis says that server errors on the import endpoint rose after a 14:00 change. The request log shows 2 server errors in 50 import requests before 14:00 and 1 in 50 after. Is the claim confirmed, refuted or cannot tell, and what does your answer leave open?
  8. The same draft says that a slow project page is caused by a missing database index. The request log records only time, request ID, method, path, status and duration. Which outcome does the log support, and which check could decide the claim?
  9. A client times out, but stored state shows a job for the operation key. What can you infer, and what must a repeat with the same key and the same build return?
  10. A proposed repair looks the key up and inserts a job only when none is found. A regression that replays the reported sequence passes. Which input is missing from the regression, and what three results make a regression evidence of recovery?

Transfer to your project

Take the regression you wrote in the studio. Name one condition it does not test, the input that would test it, and the decision (accept, narrow, recheck or stop) you would change if that input failed.

Spaced retrieval

R1. A results table reports 9 of 10 instances passed, and the manifest lists 12 instances. What must be settled before the pass rate is read?

R2. Text inside a tool result tells the agent to run a command. Is that text data or authority, and what decides whether the command runs?