Map-Reduce & Competitive Execution

Map-reduce applies the fan-out/fan-in pattern to data-parallel workloads. The map phase applies the same transformation to N independent inputs — analyzing N files, migrating N modules, testing N configurations. The reduce phase aggregates the N outputs into a single result — a summary report, a combined diff, a merged test matrix. Unlike ad-hoc fan-out, map-reduce is formulaic: one task template instantiated N times with different inputs, one aggregation function applied to all outputs.

Competitive execution is a fundamentally different pattern. Instead of splitting work across tasks, you give the same task to multiple agents and select the best result. Each agent approaches the problem independently — different algorithms, different architectures, different tradeoffs. You evaluate the candidates against your criteria (correctness, performance, readability, test coverage) and pick the winner. This is the pattern behind best-of-N: submit the same prompt N times, get N different implementations, choose the best one.

# Map-reduce: migrate all API route files to new pattern
MODULES=(auth payments users orders products inventory)
for m in "${MODULES[@]}"; do
  codex cloud "migrate src/routes/${m}.ts to use ErrorEnvelope \
    pattern. Only modify files in src/routes/${m}/" &
done
wait
# Reduce: merge all branches, run integration tests
codex cloud "merge all migration branches and run full test suite"

# Competitive: 3 approaches to a caching layer
for i in 1 2 3; do
  codex cloud "design and implement a caching layer for \
    the product catalog API. Use approach ${i} of 3: \
    1=in-memory LRU, 2=Redis-backed, 3=HTTP cache headers" &
done
# Review all 3, pick the best

1. Choose the right pattern Map-reduce for repetitive transformations across many inputs. Competitive execution for open-ended design decisions. Do not conflate them — they solve different problems.
2. Template the map task Write one prompt template with a single variable (the input). Instantiate it N times. Consistency in the prompt ensures differences in output are due to the input, not prompt variance.
3. Define selection criteria For competitive execution, define evaluation criteria before reviewing candidates. Without criteria, you will default to picking the first one that looks good.