CWI Scientific Meeting · 8 September 2026

Agents for
research

Less friction between a scientific question and a result you can check.

Get unstuckCheck evidenceExplore ideas

Syver Døving Agdestein · Xinyuan Wang

QR code linking to the slides and sketch workflow Scan to follow along agdestein.github.io/talks
half a year ago

Who are we?

Syver Agdestein

Pre-agentic Syver

Scientific Computing

  • Hunt for bugs in numerical solvers
  • Write job scripts and restart failed runs
  • Search old papers to check novelty
  • Redo plots after changing one parameter
Xinyuan Wang

Pre-agentic Xinyuan

Computational Imaging

  • Turn methods from papers into code
  • Tune settings and rerun reconstructions
  • Track down why an image looks wrong
  • Redraw figures and move labels by hand

The shared problemTime spent between a research question and a result we can check.

the opportunity

Three ways to move research forward

Three visual motifs: untangling a path, checking evidence with a magnifier, and branching into new ideas. Get unstuck Inherited code Failing runs Check evidence Claims ↔ results Papers ↔ hypotheses Explore ideas New experiments New explanations

Delegate a concrete task. Keep the scientific judgment.

Coral: researcher · blue: agent · green: check · yellow: hypothetical

from conversation to action

A chatbot replies. An agent acts.

It can suggest what to do. You do the work.

Illustrative comparison · text-only chat vs. an agent with access to tools.

shared context · one project, many tasks

The filesystem is the agent’s context.

A research project filesystem contains experiment plans, hypotheses, literature, code, results, research notes, and a manuscript. The agent chooses which files to read for the current task. project/ tracked by git experiments.md run plans proposal.md hypotheses lit/notes/ literature code/ analysis results/ measurements decisions.md research notes paper/ manuscript Does our latest result support the hypothesis? Agent chooses what to read Claim + evidence with file references
One example; more tasks available:

Illustrative research project · the agent searches and reads relevant files as needed.

get unstuck · simplified project example

Why is the saved model worse?

Training updates parameters from theta zero to theta two, but the checkpoint still references theta zero. Reveal the fix: save updated state and compare the reloaded model with the live model. Training updates θ₀ θ₁ θ₂ stale reference saved θ₀ Before: reload mismatch θ₂ After: reload save updated state matches live state

The agent traces the state. You check the reloaded model.

Schematic reconstruction · SymmetryCode, commit 885cf4d · θ denotes model parameters.

from a new question to a new experiment · project example

Build the tool your experiment needs.

The same schematic field is shown as a serial reference and split into four GPU partitions. Checks compare independent implementations and agreement across decompositions. Serial reference independent FFTW compare Agent-built solver 4 GPU partitions 1 2 3 4 Same partition byte-exact Repartitioned ≈10⁻¹² tolerance

A larger experiment starts with a solver you can check.

Partition schematic · validation criteria from DistributedNavierStokes · seven-week project, not a measured speedup.

bounded execution · hypothetical overnight campaign

Delegate the waiting and follow-through.

Two swimlanes separate researcher decisions from agent execution. Evening planning leads to queued runs, one agreed retry for preemption, a flag for numerical divergence, and a morning review of checks and failures. YOU AGENT Evening plan + budget Overnight run + monitor preempted: retry once ! diverged: flag Morning review checks + failures

Wake up to a record you can inspect.

Hypothetical sequence · retries follow an agreed policy; numerical failures are flagged.

the shared pattern · from examples to one complete task

Give every task a check.

Four evidence checks: compare code with a baseline, anchor literature claims in passages and conditions, quantify uncertainty and scope, reproduce review findings. Code baseline / invariant Literature passage + conditions Result uncertainty + scope Review finding reproduce the problem

Give a concrete goal. Choose how to check the result.

Checks illustrated here draw on the ProbabilisticClosure audit.

open-source reference case · Xinyuan · from request to result

A sketch is fast. An editable figure is not.

turn this sketch into a figure I can still changeresearcher goal

Follow one task from a rough sketch to a figure we can actually edit.

researcher’s sketch
→
Editable reconstruction of the public reference case, Candidate C editable reconstruction
before generation · resolve what the sketch leaves unclear

Clarify before generating.

Codex asks

Example answer

What does the feedback loop mean?

Optimize the network for one observation, using the difference between predicted and observed measurements.

What should the image panels show?

Fixed random noise as input, a synthetic mountain as the reconstruction, and schematic measurement curves.

What may change across the five candidates?

Explore different layouts and visual styles. Keep the method, labels, equations, and arrow directions consistent.

Same scientific meaning. Five visual directions.
public reference case · five saved visual candidates

Five proposals. I choose C.

Codexcoordinates the brief
→
ImageGencreates A–E
→
researcherselects direction
Actual candidate AA
Actual candidate BB
Actual candidate CC · SELECTED
Actual candidate DD
Actual candidate EE

Researcher choice

C · color
and clear module grouping

We have the look we want. But it is still just an image.
from pixels to objects · agree on the editing boundary

What do I need to edit?

The approved 8-region map of Candidate Cthe map we approved together

Edit individual objects

11 separate network layers
plus labels, arrows and curves.

Keep as replaceable images

The noise texture and
synthetic mountain image.

Keep the formula source

LaTeX stays available
when the equation changes.

Codex rebuilds the chosen design as editable objects — including 11 separate network layers.

during reconstruction · appearance and meaning

Keep the appearance. Check the science.

From candidate C

Keep the chosen look

Layout · colors · spacing

Selected candidate C supplies layout, colors and spacing
From the agreed brief

Rebuild and check

Exact labels

Correct equations

Arrow directions

The generated image guides appearance.
The brief defines the scientific meaning.

Preserve the visual choice without losing the scientific relationships.

try it · the result is an editable artifact

Now I can actually change it.

Drag any object
Deep Image Prior Candidate C fidelity v2 review Editable mixed-media composition with native gradients, reference-fitted synthetic plot paths, colored vector equations, and two replaceable raster atoms. Final scientific approval is pending. {"figure_id": "deep-image-prior-c-fidelity-v2", "status": "AWAITING_FINAL_RESEARCHER_REVIEW", "approved_raster_atom_count": 2, "validation_scope": "Programmable provenance, exact crop pixels, local synthetic curve-fit error, vector structure, equation aspect ratio, cross-format parse validity, and topology only; not scientific correctness, publication approval, or pixel editability of the two image atoms."} optimize θ
Drag objects to rearrange · arrows move independently.
Download SVG to save; reloading restores the original.
finish the task · separate file checks from scientific judgment

Validate the files. Then ask the researcher.

Does the file work?

Opens correctly
Parts move separately
Appearance matches
File checks and visual comparison

Is the figure right?

Correct labels
Correct arrows
Correct scientific meaning
The researcher makes the final judgment

Spend more time on your ideas,
less on drawing tools.

from these examples to your own work

Your next research task.

For your own research

What would you ask an AI agent to do?
How would you check the result?

QR code linking to the slides and sketch workflow Slides & sketch workflow agdestein.github.io/talks
check evidence

Can we trace this claim to its source?

A manuscript and a generated table can disagree. Neither is automatically right.

Paper

The claim

What exactly does
the sentence assert?

→
Results

The reported value

Which table, run
and metric support it?

→
Calculation

The evidence

Reproduce the value.
Check its scope.

The useful output is a traceable correction, with the evidence attached.

Workflow illustration inspired by Symmetry/Notes/Review4.md · a worked paragraph example is available in the backups.

explore ideas · hypothetical literature example

Find the paper that changes the plan.

A search through papers finds a hypothetical plateau in error as resolution increases. The expected decreasing curve is dashed; the counterexample plateaus. This changes the next experiment to a noise-sensitivity test. Search for exceptions “Finer grids always help” resolution → error plateau schematic counterexample Test noise sensitivity Read the source first expected

A counterexample can change your next experiment.

Invented curves and finding · workflow inspired by ProbabilisticClosure’s literature review.

your turn to judge the evidence · live demo · hypothetical data

Check this results paragraph.

Compare notes/demo/paragraph.md with notes/demo/results.csv.
For each claim, show the evidence or what is missing.

Draft paragraph

Mean error: 6.5%

A has lower mean error than B.

A wins on every seed.

Supplied aggregate results

MethodMean errorSeeds
A0.06910
B0.07510

Watch what it reads. Judge what it concludes.

Show backup answer · all demo inputs are local text files.

backup · hypothetical demo answer

Two claims need attention.

ClaimVerdictEvidence
Mean error is 6.5%Contradicted0.069 × 100 = 6.9%
A has lower mean errorSupported0.069 < 0.075
A wins on every seedNot establishedNo per-seed results supplied

Suggested revision: “Mean relative error was 6.9% for A and 7.5% for B, each over ten seeds.”

Aggregate means do not establish per-seed wins or statistical significance. Return to demo · Discussion

ambitious goals · ordinary language

Prompt zoo

You

Review my paper as a critical referee. Write down the issues that matter before submission.

Agent · illustrative reply

I wrote review.md, ranked by severity. Two claims need stronger evidence. Shall I start with the major revisions?

review.md

Give it a goal; let it work out the steps.

Illustrative prompts and replies · each task assumes access to the relevant project files and tools.

backup · simplified legacy-code example

Turn unfamiliar code into a mental picture.

A schematic code excerpt from a force-spreading routine is mapped to a surface marker distributing force to neighboring grid nodes. This is a two-dimensional slice of a three-dimensional stencil. Inherited code void MASSWEIGHING(...) { coef[a][b][c] = ...; fnorm = ...; ... } explain marker force → grid nodes

Check the explanation against the method and a known case.

GLS3D_C / MASSWEIGHING · conceptual 2D slice; the 3D routine distributes force to eight nodes.

backup · growing project context

Give the next task a useful starting point.

One shared project folder

project/ ├─ instructions.md ├─ code/ ├─ data/ ├─ paper/ └─ notes/

Short instructions

What the project is about

Where the inputs and outputs live

How to run and check the work

What requires your review

Record useful decisions as the project grows.

Illustrative layout · explicitly ask your tool to read instructions.md; automatic loading is tool-specific.

backup · communicate · interactive illustration

Let your audience explore the idea.

A schematic static paper figure becomes an interactive sine-wave plot. Move the frequency slider to explore how parameter k changes the curve. PAPER agent EXPLORE THE SAME IDEA y = sin(kx)
1.0

Move a parameter. See what changes.

Teaching illustration: y = sin(kx) · inspired by interactive research posts; not project data.

Agents for research · backups