I started building agent fleets in 2025. In 2026, the work became an operating discipline: coordinate the agents, inspect their work, and ship systems people can use.
A contribution graph shows the rhythm of the work. The systems below show what came out of it.
The first agent workflows. Learning how to divide work, coordinate agents, and keep the operator in the loop.
January 1–December 31 · full calendar year| Month | Contributions |
|---|---|
| Jan | 0 |
| Feb | 0 |
| Mar | 76 |
| Apr | 405 |
| May | 72 |
| Jun | 212 |
| Jul | 88 |
| Aug | 128 |
| Sep | 0 |
| Oct | 0 |
| Nov | 0 |
| Dec | 0 |
Turning the foundation into a repeatable practice. Shared context, deliberate review, and useful systems that reach production.
January 1–September 24 · year to date| Month | Contributions |
|---|---|
| Jan | 0 |
| Feb | 1,600 |
| Mar | 3,724 |
| Apr | 2,417 |
| May | 1,696 |
| Jun | 4,154 |
| Jul | 2,801 |
| Aug | 829 |
| Sep 1–24 | 541 |
Both calendars use the same daily-count color scale. Outlined cells fall outside the reporting period. The periods differ; these totals are not a year-over-year productivity comparison.
Download the daily activity data ↓Open a product. Inspect a public repository. Read the architecture. Each gives you a different way to evaluate the work.
Give an idea a useful next step.
Blueprint assesses an idea and turns it into an execution blueprint. Its public leaderboard makes example assessments visible before you submit your own.
A standalone coding-agent runner for routing tasks through configured agent CLIs. Public source gives you something concrete to inspect alongside the contribution counts.
Explore the public repository ↗Shared memory for the fleet: decisions, lessons, and project context assembled at the start of an agent session. The published case study explains the design; its implementation stays private.
Read the memory-system case study →Start with the source label. Public records, private historical reports and unverified estimates answer different questions. Expand a receipt for its scope and limitations.
6 receipts · expand a row for its source and limits
Daily contribution counts fetched from the signed-out GitHub calendars. Private repository activity is anonymized. These counts measure activity, not code quality or human-equivalent productivity.
Inspect the GitHub profile ↗Owner-reported count of agent launches from a Claude Code Insights report covering a 20-day window in March 2026. This is not a count of unique agents or agents running simultaneously. The report and its session logs are private.
The March 2026 report is not published, so there is no public source to link.
Owner-reported session, active-repository, and commit totals from the same March 2026 report. A commit does not establish that a change was merged or deployed. These totals cover a different period and scope from the public GitHub calendars.
The March 2026 report is not published, so there is no public source to link.
Recorded in March 2026 as 1,060 hours of engineering output across the 20-day window (about 53 per day). How the report calculated those hours is not documented, so this is an unverified historical estimate. It does not establish measured agent runtime, elapsed human time, time saved, or team-size equivalence.
The March 2026 report is not published, so there is no public source to link.
The March 2026 report records 84 security findings fixed. Finding-level records remain private, so this is an owner-reported historical result, not an independent security certification.
The March 2026 report is not published, so there is no public source to link.
The March 2026 report recorded buggy-code events and mid-session changes of approach. These show the need for review and intervention; this page does not independently establish resolution of every event.
The March 2026 report is not published, so there is no public source to link.
A conviction calculation was suppressing useful signals.
The InDecision Engine case study describes a conviction ceiling bug in its scoring formula.
The formula was corrected so signals the previous calculation missed could surface.
Inspect the decision logic, not only whether the system runs. More activity alone would not fix the formula.
Owner-authored case study. The underlying incident records are private.
The hivemind-crypto GitHub contribution calendars were fetched without signing in. The snapshot covers all of 2025 and January 1–September 24, 2026, including the private activity GitHub exposes as anonymized counts. This is a dated snapshot, not a live feed.
The daily counts sum to the displayed totals. Contributions can include several kinds of GitHub activity; they are not interchangeable with commits, hours, or shipped outcomes.
How GitHub counts contributions ↗The March 2026 figures come from a private Claude Code Insights report covering a 20-day window. That report is not published, and the figures are owner-reported, not independently audited on this page. How it calculated the 1,060 hours is not documented, so that figure is shown as an unverified estimate.
The original coverage and task-completion percentages are omitted here because their denominators and repository scope are not published. The velocity report is a separate, later analysis with different windows and metrics; it is not the source of these figures.
Read the separate velocity report →The 2025–2026 fleet narrative is Jeremy’s account of building and operating the systems. GitHub substantiates contribution activity. Product links and case studies supply the examples; activity totals alone do not prove mastery or quality.
Define a bounded task, inspect the output, and decide what happens next. Start with a workflow you can understand before coordinating a fleet.
