What does an agentic outcome actually cost?
The field begins with a practical question — once verification, recovery, inherited context, and downstream consequences are included — and it asks that question as tasks, environments, workflows, and time change.
Agentic Dynamics is the empirical study of how agents behave as tasks, environments, workflows, and time change. It begins with a practical question: what does an agentic outcome cost once verification, recovery, inherited context, and downstream consequences are included? The field's method is to instrument real agent work, derive named information from the resulting record, test policy choices as controlled arms, and repeat only where uncertainty remains. This site publishes the instrument, the corpus boundary, the verdicts, the nulls, and the open problems. It does not sell certainty, a routing product, or a universal best practice. Positioning statement — binding editorial decision
The question is deliberately not a price sheet. A per-prompt price does not say what a verified, recovered, inherited outcome costs. Answering the question requires measuring accepted outcomes under controlled change — and that requirement is what makes an instrument necessary.
One comparison became a cumulative problem.
A clean-versus-degraded prompt comparison can answer one bounded question. Agentic software work is cumulative: each session changes the codebase the next session inherits. N linked sessions are required because the output of one session is the starting point of the next. M measurement angles are required because passing tests cannot stand in for verification depth, maintainability, cache economics, recovery, or cost per accepted outcome.
Collection is only half the multiplication. The same sessions must then be analyzed across models, conditions, tasks, tiers, commits, and story positions — a second N × M burden. More sessions without analysis produce volume, not information; more metrics without consistent event capture produce incomparable claims.
What remains untested.
The field publishes its open problems as named states, never as zeros or stale fallbacks. The current open list is short and specific.
Each open problem names the signal it needs: a triggered escalation cell, a live LSP series, a powered calibration run, a canonical lab output. When the instrument measures them, they move from null to evidence.