Execution-layer benchmarks

Every agent iteration depends on execution.

An agent needs to run the application, inspect the result, and use that evidence to make its next change. Vitron is building infrastructure for that loop. This page examines two costs within it: execution latency and the context consumed by verification.

Infrastructure for repeated verification

Measure the work inside the loop.

Every verification run has a cost. Execution is a recurring part of agent-driven development. Each run takes time and returns information the agent must process. Measuring those costs helps teams evaluate the execution layer they build on.

Time to a result

Tool-call latency contributes to how long an agent waits while exercising an application.

Context per workflow

Returned tool output consumes context as the agent reviews a run and plans its next step.

Evidence for a decision

Speed is one consideration. Policy controls and reviewable run evidence also matter when choosing execution infrastructure.

Explore the platform →

Runtime latency

Tool-call latency

Reported timings for selected scenarios in Alethia’s browser execution layer, with Playwright-based tools as reference baselines.

Alethia~13 msreported typical call

Playwright MCP~200 msreported typical call

Playwright CLI~2 sreported call scenario

Call timing and flow timing are different. The published materials also report 2–5× faster flows versus Playwright MCP and up to 50× versus the CLI path. Consult the methodology for those scenarios; these figures do not measure the entire development loop or verification quality.

Agent context

Reported token counts per verification flow

Each number is the reported token count for one complete workflow. For example, the CRUD flow reports 376 tokens for Alethia and 3,175 tokens for the Playwright MCP baseline. See the methodology for counting details.

Workflow	Playwright MCP tokens	Alethia tokens	Reported token reduction
CRUD	3,175	376	8.5× fewer
Search	3,617	538	6.7× fewer
Sign-in	810	304	2.7× fewer

The existing cost illustration uses Claude Sonnet input pricing at $3 per million tokens for 1,000 CRUD flows. Model pricing changes; treat that estimate as an example, not a current quote.

Reproducibility

Run the proof yourself.

Alethia Anvil is our public reference app and reproduction kit. Run the supplied scenarios locally or adapt them to your own application.

Reference appA static application with sign-in, task management, and search.
14 test specsPlain-English workflows, a safety-policy check, and a CI runner.
Benchmark kitThree shared flows, saved results, and a documented baseline.

Alethia Anvil on GitHub Read the methodology

Results vary by environment and version.