Part 5 of 6 · 2 min read
Exercises
Easy
Easy 1: Run it and read the tables (30 minutes)
In code/mini-fanout, run pytest -q and python run_all.py.
Done when: you report the latency ratio at density 0 and 1.0, and say why the split can be slower than one agent.
Hint
fanout.compare(params, n, density) returns every number in the tables.
Easy 2: Count levels and edges (30 minutes)
Write three dependency lists by hand (independent, a chain, a diamond), then check them with fanout.levels.
Done when: you give the levels and edges of each and say which would benefit from splitting.
Hint
fanout.levels([[], [0], [1]]).
Easy 3: Price your own task (30 minutes)
Estimate your own system prompt, background and step sizes and put them in Params.
Done when: you report tokens, latency and success for one single agent and one split, and say which parameter you are least sure of.
Hint
replace(Params(), system=..., shared=...).
Medium
Medium 1: Find the break-even shared background (2 hours)
Sweep shared from 0 to 30,000 at two cache prices.
Done when: you plot token ratio against background size and report the crossover for each price.
Hint
compare(replace(P, shared=s, read=0.1), 8, 0.1).
Medium 2: A better handoff (2 hours)
Sweep handoff_loss from 0 to 0.2 at density 0.5 for both hazard settings.
Done when: you report where splitting stops being more reliable and say what a typed handoff would be worth.
Hint
replace(P, handoff_loss=h, ctx_err=0.0001, tool_err=0.0001) for the robust setting.
Medium 3: Uneven sub-tasks (3 hours)
Give sub-tasks different step counts and compute latency from the longest chain by steps, not by levels.
Done when: you show that one slow sub-task can dominate the latency and say how you would rebalance.
Hint
replace levels with a longest-path sum of step counts.
Hard
Hard 1: Real tasks (5 hours)
Run ten labelled tasks (five parallel, five coupled) as a single agent and as a split design. (I did not run this, so your numbers are the ground truth.)
Done when: you report accuracy, total tokens, latency, the model and the date, and say whether the model's ranking held.
Hint
label each task before running it so the labels cannot be influenced by the results.
Hard 2: Calibrate the hazard rates (4 hours)
Measure a real agent's per-step error against context length and tool count, and fit ctx_err and tool_err. (I did not run this.)
Done when: you report the fitted values and say whether splitting is predicted to help at your task's size.
Hint
regress failure rate on context size in 1,000-token buckets.
Hard 3: Add a critic (4 hours)
Add a reviewer agent that reads every result and rejects wrong ones. (I did not run this.)
Done when: you report the extra tokens and latency, the reliability gain under both hazard settings, and say when the critic pays for itself.
Hint
a critic adds one call per sub-task that reads the result.