old harnesspaused < 2 hours
68,000 commits
The swarm created 54 crates and three competing database packages. A total of 1,173 agents edited the most contested file and produced 7,771 conflicts.
commit activity
Attention Deficit · weekly note · serial AD-W30-7F2
Cursor experiment · July 20, 2026
Agent swarms · the second attempt
The swarm isn’t the speed.
Cursor sent fleets of agents after SQLite. The second run worked because each agent had less context to hold.
Switch the system at right. The old swarm confuses activity with progress. The redesign separates planning from implementation.
merge conflicts<1,000 / 4 hours
final shape9 crates
held-out grade80% after 4 hours
Narrow workers execute. Planners keep architecture in view. Shared notes carry decisions across the fleet.
01 · activity ≠ progress
68,000 commits produced 70,000 merge conflicts
The old Grok run produced 68,000 commits in under two hours and more than 70,000 merge conflicts. Activity outran coordination.
new harnessfull 4-hour run
Fewer than 1,000 conflicts
The second run settled on nine crates early. Its most contested file saw 47 conflicts, and the database scored 80% on held-out tests after four hours.
commit activityCoordination catches conflicting decisions before they reach the merge queue.
02 · swarm repair manual
At machine tempo, ordinary teamwork breaks.
Cursor rebuilt ownership, review, merge queues, and shared decisions for a fleet peaking near 1,000 commits per second.
failure 01 · two realities
Two planners invent the same idea differently.
The codebase develops incompatible definitions of the same concept before either planner sees the other branch.
REPAIR → Keep design decisions with planners, then partition tasks so two planners don't decide the same question.
03 · model routing
Most tokens don't need your smartest model
Planners remove ambiguity. Once the instruction is explicit, cheaper workers can process most of the tokens without changing the destination.
new system · similar end quality
frontier model · planner + workers
$10,565
baseline · one frontier model everywhere
In the all-frontier run, workers cost $9,373 because execution processed most of the tokens. Only a few decisions required frontier judgment.
04 · the abstraction shift
The unit of work becomes the spec.
Autocomplete raised the unit from character to line. Agents raised it from file to feature. A swarm turns an explicit description of intent into an entire task tree.
source language
“Implement SQLite from its manual.”
→
planner instructions
Architecture + bounded task tree
worker instructions
Narrow, explicit tasks
→
machine output
Code, tests, reviews, reconciled decisions
runtime checks
Compiler + held-out database tests
→
scarce input
A precise description of intent
SEMANTIC CHECK → A compiler preserves meaning at every stage. A swarm can change it at every stage, so the coordination system checks each handoff.