Merge/review time as a fraction of total session time grows superlinearly as the number of parallel Dakka agents increases beyond 4, confirming the Amdahl's Law prediction from the agent-scaling research

Hypothesis

Merge/review time as a fraction of total session time grows superlinearly as the number of parallel Dakka agents increases beyond 4, confirming the Amdahl's Law prediction from the agent-scaling research

Changelog

Date	Summary
2026-04-06	Audited: added Changelog, domain tag, stamped last_audited
2026-04-04	Initial creation

---

Hypothesis

The agent scaling research predicts that review/merge time becomes the dominant bottleneck as parallel agent count increases. At ~10 parallel sessions, merge conflicts become a “full-time job.” Dakka is a parallel agent orchestrator that spawns multiple Claude Code sessions. The hypothesis is that this Amdahl’s Law effect is measurable in Dakka: the serial fraction (merge + review) grows from <10% at 2 agents to >40% at 10 agents, making additional parallelism counterproductive.

Method

1. Controlled task: select a well-defined codebase modification task (e.g., “add unit tests to 10 modules” or “migrate 10 components from library A to library B”). The task must be decomposable into N independent subtasks
2. Parallelism levels: run the task 5 times with different agent counts: 2, 4, 6, 8, 10 agents
3. Instrumentation: use BloomNet session logs to measure:
   • Agent work time: total time agents spend actively generating code
   • Merge time: time spent resolving git conflicts between agent outputs
   • Review time: time human spends approving/rejecting agent outputs
   • Idle time: agents waiting for merge slot or human approval
4. Serial fraction computation: S = (merge_time + review_time) / total_wall_clock_time at each parallelism level
5. Amdahl’s curve fit: plot S vs N (agent count). If the relationship is superlinear, it confirms the O(n^2) merge complexity prediction
6. Optimal parallelism: find the N where total wall-clock time is minimized (diminishing returns threshold)

Results

Pending. Will measure:

• Serial fraction at each parallelism level
• Total wall-clock time vs parallelism level
• Merge conflict count vs parallelism level
• Optimal agent count for the test task

Findings

Pending.

Next Steps

If the bottleneck is confirmed, implement mitigation strategies from the research: (1) pre-partitioned file ownership per agent (eliminates merge conflicts by construction), (2) automated merge validation (agents resolve their own conflicts before human review), (3) progressive merge (merge pairs of agents, then merge merged results, reducing O(n^2) to O(n log n)).