Measured execution. Estimated counterfactual. One operator.
This dataset records what agentic systems do for me, not what they will do for the average engineer. It is an upper-tail, n=1 operational record built in public and published so the arithmetic can be checked.
I bring more than 35 years of professional experience across software engineering, enterprise architecture, and infrastructure architecture; strong product judgment; direct authority to decide what gets built; an extreme seven-day work cadence; and multiple agent systems operating in parallel.
The interaction among those conditions is the point. Agentic engineering lets me follow ideas from conception through production without organizational handoffs. The practical limits are increasingly judgment, mental bandwidth, and sleep. This record should not be generalized into a staffing formula or a promise about an average team.
The checked-in publication snapshot is the authority for these figures. The live Fulcrum service is not used to refresh this page after publication.
Dates, agent-session duration, and token counts when provider telemetry is available.
Qualified-senior hours and tokens when exact telemetry is unavailable. Estimates remain explicitly labeled.
Execution factors, medians, means, and rollups recomputed from the published source fields.
My judgment about whether work was feasible unassisted, required substantial study, or expanded my capabilities.
Senior-equivalent minutes divided by total agent-session minutes across the snapshot. This ratio of sums describes the body of work without giving small, high-factor tasks disproportionate weight.
Σ senior minutes / Σ agent minutes The median describes a typical task. The arithmetic task mean describes the average of individual task factors. Both are distribution statistics, not the aggregate execution factor.
Senior-equivalent minutes divided by my prompt, intervention, correction, and review time. It measures the return on the resource I personally supply.
Senior-equivalent work compared with elapsed clock time. It captures concurrency, unlike a sum of individual session durations.
Work I could do unaided, work that would require substantial learning, and work I could not reasonably complete alone are separate classes. The last class should be counted, not assigned an invented infinite multiplier.
Why a day can contain more than 24 agent-session hours. Sessions overlap. Three agents running for ten hours each produce 30 agent-session hours inside ten elapsed hours. Additive session time measures consumed execution; calendar compression measures parallel throughput.
Versioned public snapshot
The archive contains CSV and JSON records, a manifest with row counts and checksums, and methodology and correction notes. Task descriptions are redacted because the originals name internal systems and unreleased work; the quantitative fields are retained.
leverage_factor is recomputed from the published estimate and duration. Token provenance matters: exact provider telemetry and rate-based estimates must be distinguished rather than described together as measured.
Interpretation boundary
These results emerge from the combination of deep engineering and architecture experience, product judgment, an accumulated automation layer, expensive access to several frontier systems, high concurrency, direct product ownership, and a work cadence that most people would neither choose nor sustain.
That combination can push personal leverage beyond conventional expectations. It does not imply that one randomly selected engineer replaces a team, that every project should produce the same factor, or that an organization can acquire these results by purchasing subscriptions.
Some estimates are probably conservative from my personal perspective. In specialized work, including tensor and mathematical implementation in AVIAN, the relevant counterfactual may include months of study or no reasonable unaided path at all. The dataset will preserve the qualified-senior estimate for continuity while recording personal feasibility separately.
I believe many personal-unassisted durations may be four to five times the qualified-senior baseline. That is a sensitivity view, not a corrected fact. Historical rows are not receiving a blanket multiplier; personal ranges and capability classifications will be added only through reviewed evidence.
The ledger
Daily entries preserve the sequence and task-level context behind the snapshot. They are historical records; current methodology notes govern how their claims should be interpreted.