Execution, inputs and outputs, cost, errors, and history, all recorded as graph structure you can query, replay, and compare.
Monitoring is not an add-on. The runtime records itself by construction.
Every run is nodes and edges: steps, tool calls, decisions. "Why did it do that?" is a graph query, not a log grep.
Token spend, latency, and failures per agent, per flow, per day, queryable like any other data.
MVCC snapshots let you re-run a decision with a different model or prompt and diff the outcomes.
Six months later: which runs touched this customer? What did the agent know at the time? Time-travel answers both.
Ather Forge is in active development. Early-access partners get direct engineering access and shape what ships first.