Pareto Front

Define “Better” Before Picking A Model

Choose the lowest-cost model route that meets your quality and response-time requirements—not the model with the best score in isolation.

If you’re shipping a product that calls models, “best model” is an incomplete requirement. An interactive extraction workflow and an overnight batch can use the same task logic while needing different cost–latency trade-offs.

Use Constraints To Narrow The Choice

Suppose two model routes pass your extraction quality checks:

• Route A: cheaper, but slower.
• Route B: faster, but more expensive.

Neither dominates the other. Improving cost means giving up speed. That’s the useful idea behind a Pareto front: a set of trade-offs rather than one universal winner.

For an interactive workflow, choose the cheaper route that still meets your response-time limit. For an overnight batch, that limit may be much looser.

Before comparing routes, write down:

• Quality: What must the extraction get right?
• Latency: How long can the user or downstream job wait?
• Cost: What can you spend per completed task?

Then test against those constraints. A route that misses a hard requirement isn’t “better” just because it wins another metric.

Keep The Workflow Separate From The Model Decision

Model choice is a decision you should be able to revisit. Rebuilding the agent execution loop every time is not a useful part of that evaluation.

Kimss Forge, from Kimss AI, is an MIT open-source agent harness. It runs locally with your model endpoint, is free, and requires no Kimss account.

Start with:

pip install kimss-forge

Build the workflow with a few lines and your endpoint rather than spending another afternoon rebuilding the same execution loop. Forge’s BENCHMARK.md documents the comparison behind the about 80% less app code claim—that is a code comparison, not a model-cost or latency claim.

Without a production gateway, a sensitive tool invocation produces a soft Authority Boundary warning in the terminal. It never blocks local execution; a warning is not gateway enforcement.

Attach Production Controls Without Replacing The Agent

When you need controls on production model calls, add gateway="kimss" to the same Forge agent.

For an existing OpenAI-compatible client, the gateway address goes in its base_url configuration:

base_url = "https://api.kimss.ai"

That line shows the endpoint change, not a complete authenticated setup. You still need your gateway credentials and connected model infrastructure.

Kimss is a model-agnostic API gateway, not a chat app or a model host. You bring your agents and infrastructure.

On the free Developer path, gateway-routed traffic gets:

• Identity on model calls.
• Kill Switch enforcement at the gateway, including mid-hop enforcement.

The scope matters: the Kill Switch severs access through Kimss. It does not stop customer processes that never call the gateway.

Authority Boundary, SSO, and SCIM can unlock later on the same path without a rewrite; they are not free Developer features.

Make Model Choice Reversible

Start with the task and its constraints. Keep the workflow as the foundation, evaluate model routes against explicit requirements, and add gateway controls when moving toward production.

Next step: run pip install kimss-forge.

If you ship with models, put a control plane in front of them—start with Forge.