Cost per Feature

Cost per Feature is the fully loaded cost of delivering each feature through agent execution. It aggregates four cost components:

1. API token consumption — the total tokens consumed across all agent runs required to complete the feature, including retries, Rescue Mission interventions, and evaluation passes.
2. Compute infrastructure — the cost of Ephemeral Workbench provisioning, test environment runtime, and any supporting infrastructure used during execution.
3. Human supervision time — the labor cost of Context Architect time spent writing Live Specs, Agent Operator time spent on oversight and Rescue Missions, and Evaluation Engineer time spent maintaining the Eval Harness.
4. Review effort — the labor cost of human code review after agent output passes automated evaluation.

Cost per Feature is calculated per delivered feature:

Total cost (tokens + compute + supervision + review) / Number of features delivered

This metric determines whether agentic development is economically viable for a given team and work profile. If agent-delivered features consistently cost more than the equivalent human-delivered features — including the full salary, benefits, and overhead of a traditional team — then the investment is not paying off for that category of work.

Three strategies reduce Cost per Feature:

1. Long-term memory and caching — caching frequently used context, previous solutions to similar problems, and compiled dependency artifacts reduces the token and compute cost of each run. Agents that start with relevant cached context consume fewer tokens reaching a solution.
2. Task routing optimization — routing only positive-ROI tasks to agents and keeping negative-ROI tasks with human developers. Not all work benefits equally from agent execution. Routine, well-specified tasks with clear patterns tend to have low Cost per Feature, while novel or ambiguous work tends to have high Cost per Feature.
3. Model selection — using smaller, less expensive models for routine tasks and reserving larger models for complex work. A task that can be completed correctly by a smaller model should not consume the Token Budget of a frontier model.

Cost per Feature is reviewed during the monthly FinOps Review alongside Blended Efficiency and Token Budget trends. Together, these three metrics provide the economic case for agentic development at the unit level (Cost per Feature), the team level (Blended Efficiency), and the resource level (Token Budget).