About

Mission

Fizeau exists for three reasons that build on each other:

1. Facilitate agentic development. A reusable, embeddable agent loop with the right primitives — tool-calling, planning, compaction, retries, sampling, reasoning, quotas, session logs — so building tools doesn’t mean re-implementing the loop every time. Other tools embed fizeau.New(...) instead of writing their own agent harness.
2. Make agentic work measurable. Per-turn timing, prefill vs decode breakdown, cost-per-trial, subscription-quota accounting, route-attempt feedback — all first-class outputs, not bolted-on observability. You can’t improve the prompts, agents, or providers you can’t measure.
3. Make local models a real option. Local serving (vLLM, MLX, LM Studio, Ollama) on the same provider surface as cloud frontier models. The benchmarks compare them honestly. Self-hosted at the right quantization is often cheaper, sometimes faster, and rarely the right answer for everything — but you can pick per workload because the data is on the table.

Harness control from #1 enables #2 (we can instrument what we own) and makes #3 viable (one provider surface across cloud and local). Tools built on fizeau inherit that without writing their own loop, sampling controls, performance instrumentation, cost tracking, or subscription accounting.

Why “Fizeau”

Named for Armand Hippolyte Louis Fizeau (1819–1896), the French physicist who made the first terrestrial measurement of the speed of light — and then, two years later, did something more interesting.

In 1849 Fizeau spun a toothed wheel on a rooftop in Suresnes and bounced a beam of light off a mirror 8.6 km away on Montmartre. By spinning the wheel fast enough that the returning beam was blocked by the next tooth instead of the gap that emitted it, he could solve for the round-trip transit time of light over a known distance — and from that, the speed of light. He measured 313,000 km/s, within a few percent of the modern value. The point is not the result. The point is that you can pin down a quantity that fast by making the measurement chain itself fast and precise enough to interrogate it.

In 1851 he sent two coherent light beams through a tube where water flowed in opposite directions, recombined them, and looked at the interference fringes. The fringes shifted with flow direction — proving that the speed of light through a moving medium depends on the medium’s motion, but not by the simple sum a Newtonian intuition would predict. The “Fizeau drag coefficient” he derived matches the value Einstein arrived at from special relativity sixty years later.

Two ideas embedded there:

1. The medium matters. You cannot measure something travelling through a substrate without accounting for what the substrate is doing.
2. Differences are easier to measure than absolutes. Fizeau didn’t measure the absolute speed of light through water; he measured the difference between light flowing with the current and light flowing against it.

Modern LLM agents are bytes flowing through a measurement chain — provider, network, harness, container, tools, and back. What we want to know about that chain is, mostly, how is it running and where is it slow? Fizeau, the software, treats those measurements as first-class output, not as observability bolted on. Every turn produces structured timing. Every provider runs through the same surface so deltas mean what they look like they mean. The rotating-wheel motif fits: a precision chronograph wrapped around a fast-moving signal, designed to interrogate a moving medium without distorting it.

Read more