The orchestrator never leaves the chat.
The model I talk to plans, routes, and reviews. It doesn't disappear into a forty-minute solo implementation.
01
The inflection.
February 2026 is where I stopped typing code as the primary act and started running agents against a written doctrine. The method changed, not the hours.
Clean contributions per month
Jan 2025 – Nov 2026 · automated snapshots excluded · Nov 2026 partial (through Nov 4, 06:00 UTC)
Scroll the chart horizontally to explore every month, or open the data table.
Data table
| Month | Contributions |
|---|---|
| 2025-01 | 0 |
| 2025-02 | 0 |
| 2025-03 | 27 |
| 2025-04 | 0 |
| 2025-05 | 0 |
| 2025-06 | 0 |
| 2025-07 | 46 |
| 2025-08 | 1 |
| 2025-09 | 31 |
| 2025-10 | 6 |
| 2025-11 | 10 |
| 2025-12 | 8 |
| 2026-01 | 34 |
| 2026-02 | 230 |
| 2026-03 | 578 |
| 2026-04 | 460 |
| 2026-05 | 633 |
| 2026-06 | 1,177 |
| 2026-07 | 1,338 |
| 2026-08 | 1,143 |
| 2026-09 | 1,027 |
| 2026-10 | 1,331 |
| 2026-11 | 202 (partial) |
Monthly contribution rate, after vs before
82×
10.8PER MONTH · 2025
880PER MONTH · FEB–OCT 2026
(7,917 ÷ 9) ÷ (129 ÷ 12) ≈ 81.8. November is excluded while partial. The grey sliver is 2025's average drawn to the same scale.
02
Every dot is a contribution.
2025 is a small strip. 2026 builds a wall. Hover the wall to see which month each block of dots came from.
129 vs 8,153
2025 · full year129
2026 · to Nov 48,153
Alternating shades mark month boundaries. Excludes 13,094 automated Lab snapshot commits.
03
How work moves through the factory.
Six rules, written down and enforced by config. Each dot below is one change: written by one model family, reviewed by the other, verified against the running build, then shipped.
written by Claude
written by GPT
ring = reviewer from the other family
active rule
Research and implementation get fanned out.
Wide reading goes to a subagent that returns a conclusion, not a file dump. Independent work runs in parallel.
Model choice is a routing decision.
Hard architecture and reviews to the top tier, lead work to Sol, ordinary to Terra, mechanical to Luna.
One independent review, other family.
A model that didn't write it gets the intent, the diff, and the evidence. Read-only, fresh context, high effort.
Verify against the running build.
Reproduce first. Exercise the real flow in a real browser or binary. Tests are evidence, not the whole case.
“Fix this” means finished.
Implement, verify, review, push, CI green, PR ready. And the inverse: a question authorizes nothing.
04
Model choice is a routing decision.
Not a default. The hard reasoning goes to the expensive model; the mechanical volume does not. Labels are literal. Nothing gets relabeled to look better.
The routing table
Which tier does which kind of work
Task
GPT
Claude
Hardest reasoning & reviewsarchitecture, ambiguity
Astragpt-6-astra
Opusclaude-opus
Lead implementationcomplex features, debugging
Solgpt-6.1-sol
Opusclaude-opus
Ordinary parallel workexploration, routine edits
Terragpt-5.6-terra
Sonnetclaude-sonnet
Narrow mechanical volumerenames, sweeps, formatting
Lunagpt-5.6-luna
Haikuclaude-haiku
Tokens by model
Top five models · retained-log snapshot 2026-11-04
ClaudeGPT
Token data table
| Model | Tokens |
|---|---|
| GPT-5.6 Sol | 9,027,145,995 |
| Opus 5.5 | 5,589,573,763 |
| GPT-6 Astra | 3,580,525,928 |
| GPT-6.1 Sol | 2,481,843,094 |
| Opus 4.8 | 2,375,699,238 |
Share of top-five tokens by family
Claude · 7.97B (34.5%) / GPT · 15.09B (65.5%)
GPT-5.6 Sol accounts for ~31% of the 28.8B observed tokens. Retained logs from four machines; cleanup can reduce this floor.
05
By the numbers.
Every figure has its qualifier printed next to it. A number without its caveat is marketing.
82×Monthly contribution rateFeb–Oct 2026: 880/mo vs 10.8/mo in 2025. Complete months only.
8,153Contributions in 2026Jan 1 – Nov 4 (UTC). Excludes 13,094 automated snapshot commits.
3→1,290PRs opened, 2025 → 2026Opened, not merged. 2026 through Nov 4.
2→20Repos with commits87 exist on GitHub; these are the ones receiving work.
~1.84MLines changed · local 12 moFloor: ~1.59M added / ~249k deleted, 19 local repos.
28.8BAI tokens burnedBurn-log snapshot 2026-11-04, retained logs from four Fleet machines. A floor.
134Releases of one appmurmur-app v0.1.0 → v0.59.3, latest 2026-11-04.
22Personal agent skillsOn the collecting workstation, plus 5 hook events.
06
The stack under the factory.
Tools built to make the doctrine mechanical instead of aspirational, from the machines at the bottom to the products at the top. Pick a layer.
Fleet and the CPA gateway support agent-mesh, hangar, usher, and agentos. They power products including murmur-app, aperture, and fleetmap. Explore every layer in the paper dossier.
07
Twenty-three months, one curve.
The bar under each card is that month's contributions on the same scale as the July 2026 peak.
From manual work to agent orchestration
March 2025: first St. Basil’s commits. January 2026: murmur-app starts. February: agent workflows adopted. June: infrastructure expands. July: 1,338 contributions. October: 1,331 contributions. November: murmur-app reaches v0.59.3; 202 contributions so far.
Read the full timeline← drag or scroll →