Self-improving agents that simulate the experiment before you run it.

Explore hypothesis space on Superradiance, with each question simulated at the level of detail it needs.

Fig. 01 / SuperradianceThis field simulates the quantum effect superradiance, which happens when emitters that share one light field fall into phase and radiate together, so intensity grows as N², not N (Dicke, 1954). We build discovery agents that improve one another on the same principle.

N = 2 · burst ∝ N² = 4

Intelligence compounds.

/ Thesis

Superradiant is closing the scientific discovery loop with a novel representation of scientific knowledge that enables autonomous evaluation of hypotheses and verification of experimental results.

Frontier AI has revolutionized software engineering and mathematics. We are now at the beginning of a similar paradigm shift in physics, chemistry, and biology.

What does mathematics have that physics, chemistry, and biology still lack?

A grammar of symbols which enables a formal representation upon which frontier humans and agents can reliably compute.
Certainty. To unlock scalable intelligence in the sciences, we must model inherent uncertainty when we encounter it in a way that can be described and comprehended within this new grammar.

If discovering the Platonic forms which underlie our physical reality excites you, write us at hello@superradiant.ai

/ Our approach

Real experiments are slow and costly. Hand-built physics models break outside the cases they were made for. Our agents pick the right level of simulation for each question, check the result against evidence they held back, and improve one another from what they learn.

Read our approach

Better experiments.

/ The platform

One loop. Three ways in.

Superradiance is where you explore hypothesis space. State a question, let agents build and run the simulations, and see which hypotheses hold up. We host, you run.

ExploreSee the hypothesis space, not one guess at a time.

Start from a question. Superradiance lays out the hypotheses worth testing, tracks which ones have been run, and shows where the evidence points.

You bring: A question, its constraints, and any data you already have
Records: Every hypothesis, prompt, context, score, and execution log

SimulateRun the experiment at the right level of detail.

Agents choose how much of the world to model for each question, from a quick approximation to a detailed physical simulation, and run it on hosted compute.

Built on: A sandboxed runtime, capability allow-lists, and an explicit threat model
Already in use: LabSpine checks real lab-robot protocols in a simulator before they run

See LabSpine

ImproveAgents that get better at the next question.

One agent revises another, the revised agent runs the task, and held-back evidence decides whether the change is kept.

Mechanism: A meta-agent writes and revises a target-agent; an evaluator reads the trajectory and proposes the next generation
Guardrail: A change is kept only if it improves on cases the agent never trained on

Harnesses and learning

/ A worked example

Follow the evidence. Change the next run.

Every hypothesis on Superradiance faces the same test. Here is the smallest version, run in your browser: define a task, run a baseline, score it on training and held-out points, and keep a change only if it holds up on both.

Signal denoisingBrowser calculation · No AI model

Step 1 of 4 / Define

Define the task and keep some points aside.

Can a local average reduce noise in a measured signal? Seven training points are what we run and tune on. Seven other points are held out, and used only to check the result.

Training: measured: 2, 4, 3, 5, 8, 6, 9
Training: reference: 2, 3, 4, 5, 6, 7, 8
Held-out: measured: 4, 2, 6, 5, 9, 7, 8
Held-out: reference: 3, 4, 5, 6, 7, 8, 9

Step 2 of 4 / Run

Run a baseline.

The first method returns each measurement unchanged. Its output is the starting point that any proposed change has to beat.

output[i] = measured[i]

Baseline output (training)

2, 4, 3, 5, 8, 6, 9

Step 3 of 4 / Evaluate

Score it on both sets.

Compare each output with its reference, square the differences, and take the mean. Lower is better. The training error is measured on points we tune against; the held-out error is measured on points we never touch.

error = mean((output − reference)²)

Baseline training error: 1.143
Baseline held-out error: 1.857

Step 4 of 4 / Improve

Propose a change. Keep it only if it holds up.

The proposal averages each measurement with its neighbors; at the two edges it averages two points. The rule: accept it only if the held-out error drops as well as the training error.

output[i] = mean(measured[max(0,i−1):i+2])

Candidate output (training): 3, 3, 4, 5.333, 6.333, 7.667, 7.5
Training error: 1.143 → 0.274
Held-out error: 1.857 → 0.448
Decision: Accepted: both errors dropped

For contrast, a shortcut that copies the training answers scores 0.000 on training but 2.857 on held-out, worse than the baseline. The rule rejects it. Two small sets still cannot show that the neighbor average generalizes; that would take more signals.

Training: measured / referencen = 07

ReferenceMeasured signal

Training errorPoints we run and tune on1.143

Held-out errorPoints kept aside1.857

7 training points · 7 held-out points kept aside

An illustrative experiment, calculated from the values above. These are not Superradiant product benchmarks.

Download the record

/ What we are after

Show the work. Question the score.

The question behind the work

Superradiant is a research lab. We study how populations of agents improve one another, how to capture the signal driving that loop, and how to keep the result faithful, auditable, and safe to run.

Technical foundations

How an agent’s harness evolves

The task, execution record, and feedback loop in the self-improving agent harness, with links to the upstream work.

Read the note Research direction

Better at the task—or just the test?

Why held-out evaluation matters, and the question we want to investigate.

Explore the question

Already shipped

LabSpine.ai

Our first step toward simulating the experiment. LabSpine launched at AGI House × Google DeepMind on May 30, 2026. It turns an assay description into a robot workcell and a real Opentrons protocol, checks it in a simulator before anything runs, repairs its own errors, and traces samples across robots, LIMS, and bioinformatics. Its core is open source; Superradiant provides the hosted version.

labspine.ai Source on GitHub

/ The team