Pensero Agent Traffic Control

Rational model usage
for coding agents

Measure the cost per delivered point for each owner. Pensero keeps premium models where they deliver and routes the others to lower-cost models.

illustrative data

Measure cost per point.

Compare each owner’s agent spend with the delivery points their sessions produced, in USD per point.

Keep premium where it delivers.

Owners within target pass unchanged to the model they asked for. Low delivery is flagged, never blocked.

Route the rest.

When an owner’s $/pt sits above target, premium requests go to a lower-cost model at your provider, with reasoning effort capped.

What your agent spend delivers

See where agent spend goes

Compare session cost by output type before you change how a team works.

session cost by output type. $246 · 11 sessions.

Keep the model where it earns its cost

Compare cost per point per owner to decide where an expensive model is worth keeping.

cost per delivered point. $20 / pt default target.

See the work behind the points

Delivery points multiply the magnitude of each PR by its complexity. Our proprietary technology combines AI, heuristics and ML to give deterministic results.

pr size × code-level complexity. 6 scored PR sessions · 14 points.

See which requests changed model

Route premium requests to a lower-cost model; investigate low delivery without blocking models.

recent model requests. Policy applied to each request.

Impact-based routing

Compare the cost of token usage with the tasks that sessions delivered. Use USD per delivery point ($/pt) for each owner over a configurable 14-day window.

target $20 / pt · 14-day window
01 · Incoming request

Owner + model

Every request carries the owner it belongs to and the model it asked for.

02 · Owner rating

Check the rating system

Fewer than 3 scored sessions, or no rating? The request passes unchanged.

03 · Owner state

Check cost vs delivery

$/pt against your target. Low delivery passes, flagged only.

04 · Requested model

Premium + restricted?

Yes: fall back to a lower-cost model. Otherwise, pass.

Runs in your VPC

Your session data stays on your host. Send agent requests to any OpenAI-compatible endpoint or LiteLLM provider, including OpenAI, Gemini, vLLM and Ollama, or Bedrock or Anthropic.

one docker compose deployment

Records stay with you.

Proxy, scorer, session database and dashboard run on your host. Session data never leaves it.

Any provider you already use.

Only agent requests and their context go out, to any OpenAI-compatible endpoint or LiteLLM provider.

Score PRs in the cloud or locally.

One Docker Compose deployment, a proxy key and a password-protected dashboard. Add a local PR scoring model if you want.

Become a partner →

Questions about agent traffic control

Why measure coding agent usage?

Agent spend grows fast, and a token bill does not show what the work delivered. Pensero puts the cost of each session next to the work it produced, so you can see which usage pays off.

Which coding agents and models does it support?

Pensero scores Claude Code and Codex CLI sessions today. Agents can use any model provider, such as OpenAI, Anthropic, Gemini, Bedrock or a local model.

Do delivery points measure business impact?

They are a proxy for it. Pensero measures the magnitude and difficulty of the work that each session delivers. Use them together with a review of the work, not to rank people.

How does impact-based routing work?

Pensero compares each owner's spend with the work they delivered. Where a premium model delivers, it stays. Where the cost per point is too high, requests go to a lower-cost model. Low delivery gets a flag, not a block. You set the thresholds.

Does our data stay in our environment?

Pensero runs on your host or VPC. Session data, scores and the dashboard stay there. Model requests go only to the providers that you choose, and these can be local.

How do we deploy it?

Book a call. We deploy it on-prem, on your host or VPC.