Home › Blog › AI Doing R&D
Practical RSI · Applied Lens · Snapshot: September 2026 🔬

AI Doing R&D: From Research Assistant to Autonomous Research Organization

In 2026 AI systems were credited in Nature with discovering drug candidates, writing expert-level scientific software and generating testable biomedical hypotheses; an AI-written paper passed workshop peer review; nine Claude agents out-researched human alignment researchers on a well-defined problem; and OpenAI said ten thousand agents produced a Lean-verified Navier–Stokes blow-up proof that mathematicians are still checking. This visual guide maps the research cycle, where AI already takes part, eight landmark systems and what each actually proved, an eight-level autonomy ladder, the architecture of an AI research organization, and why research taste and verification — not agent count — now set the pace, with a runnable model to show it.

How to read this: a companion to Are We Approaching RSI? and the RSI series' AI Automating AI Research. Results are as reported by each team; we mark which are peer-reviewed, which are independently replicated and which are neither. Sources in References. The discovery-funnel simulator is illustrative.

01 The Scientific Research Cycle

Research is a loop: a goal, a literature review, hypotheses, experiment design, implementation, experiments, analysis, verification and peer review, and finally knowledge that informs the next question. Every AI research system automates some arc of this loop. Press play to walk it; the colour of each stage shows how much AI can do there today:

▶ The research cycle Colour = AI maturity at each stage, Sep 2026

02 Where AI Already Participates

AI now contributes to almost every stage — but at very different depths. The map below shows which landmark system has shown each capability. Hover a cell:

▶ Stage × system map ● demonstrated · ◐ partial · blank = not shown
Hover a cell.

Two gaps stand out. Physical experimentation is still done mostly by humans (Robin's discoveries were tested by human lab scientists). And choosing the research goal — the first column — is where no system acts on its own: every one of them starts from a problem people defined.

03 Eight Landmark Systems

Pick a system for its architecture, what it achieved, and how independently that has been checked:

▶ Landmark AI research systems Public examples, 2025–2026

A pattern runs through all eight: they work where feedback is objective. Code that runs faster (AlphaEvolve), a benchmark score (ERA), a formal proof (OpenAI's Lean formalisation), a measurable gap (Anthropic's AAR), a cell assay (Co-Scientist, Robin). Where the judge is fuzzy — "is this paper interesting?" — results are weaker: The AI Scientist-v2 got one of three papers over a workshop's bar.

04 The R&D Autonomy Ladder

From assistant to autonomous organisation there are eight levels, R0–R7. Drag the slider to climb; the bar shows who does each stage of the cycle at that level:

▶ R0 → R7 Human share vs AI share per stage

R0–R1 are widely deployed; R2–R3 are demonstrated (ERA, AlphaEvolve, AAR, Co-Scientist); R4–R5 are early and partial (Robin, The AI Scientist-v2, OpenAI's math system); R6 and R7 — an autonomous organisation that picks its own directions, and one that improves its own models and harnesses — are not publicly demonstrated. R7 is where AI R&D becomes the recursive self-improvement covered in the previous article.

05 Architecture of an AI Research Organization

The systems above share a recognisable architecture — a harness scaled up to an organisation. Humans set the agenda; a supervisor agent decomposes goals and allocates compute; specialised agent teams work in parallel on shared execution environments; results flow to verification; everything is written to a shared research memory. Press play to follow one research question:

▶ One question through the organisation Click a component for details
Click a component.

Each component maps to a real design already in use: Co-Scientist's supervisor and tournament of specialised agents; the AAR's shared forum and code store; OpenAI's thousands of parallel agents exchanging messages; ERA and AlphaEvolve's evaluator-driven search; Robin's hand-off to a human-run wet lab. What no deployed system yet has is the top box running itself.

06 Taste and Verification: the Real Bottleneck

If agents are cheap, why not run a million? Because ideas must be tested, and testing capacity is finite. What matters then is research taste — how well the system ranks its own ideas — and how many it can verify. In this model each agent proposes 20 ideas a day, 2.3% of ideas are real discoveries, and only the top-ranked ideas get tested:

▶ Discovery funnel Same model as the script below (20 simulated days per point; the script uses 50)
"""AI research organisation: are more agents, or better taste, the bottleneck?

Each agent proposes 20 ideas a day. Each idea has a hidden true value v ~ N(0, 1); an idea is a real
discovery if v > 2 (about 2.3% of ideas). Agents rank ideas with "research taste": their score is
correlated with v at strength r. Verification (experiments, peer review, formal checks) can test only
the top 50 ideas per day. We count real discoveries confirmed per day, averaged over 50 days.
Illustrative model, standard library only.
"""
import random, math

def day(agents, taste, capacity, rng):
    ideas = []
    for _ in range(agents * 20):
        v = rng.gauss(0, 1)
        score = taste * v + math.sqrt(1 - taste ** 2) * rng.gauss(0, 1)
        ideas.append((score, v))
    ideas.sort(reverse=True)
    return sum(1 for _, v in ideas[:capacity] if v > 2)

def rate(agents, taste, capacity=50, days=50):
    rng = random.Random(agents * 1000 + int(taste * 100) + capacity)
    return sum(day(agents, taste, capacity, rng) for _ in range(days)) / days

print("Confirmed discoveries per day (verification capacity: 50 ideas/day)\n")
print(f"   {'agents':>6} " + "".join(f"{'taste r=' + str(t):>13}" for t in (0.2, 0.5, 0.8)))
for a in [1, 10, 100, 1000]:
    print(f"   {a:>6} " + "".join(f"{rate(a, t):>13.1f}" for t in (0.2, 0.5, 0.8)))
print("\nWith 1,000 agents and taste 0.5, doubling verification capacity to 100/day:",
      f"{rate(1000, 0.5, 100):.1f} per day")
print("\nMore agents help only while verification can keep up; after that, taste (which ideas to test)")
print("and verification capacity set the pace.")

Output:

Confirmed discoveries per day (verification capacity: 50 ideas/day)

   agents   taste r=0.2  taste r=0.5  taste r=0.8
        1           0.4          0.4          0.4
       10           1.9          3.1          4.1
      100           3.3          8.5         20.7
     1000           4.1         14.4         38.6

With 1,000 agents and taste 0.5, doubling verification capacity to 100/day: 27.1 per day

More agents help only while verification can keep up; after that, taste (which ideas to test)
and verification capacity set the pace.

With poor taste (r = 0.2), going from 100 to 1,000 agents adds almost nothing: verification is spent on ideas that only look good. With strong taste (0.8), the same thousand agents yield nearly ten times as many confirmed discoveries. And doubling verification capacity helps more than multiplying agents once the funnel is full. This is why the frontier is shifting from generating hypotheses — now abundant — to trustworthy validation and scientific judgment, both scarce.

07 Bottlenecks and Open Challenges

▶ What still limits autonomous R&D Click a bottleneck
Click a bottleneck.

08 How Solid Is Each Claim?

Headlines flatten very different levels of evidence. Here is each landmark result by how far it has been checked:

ClaimPeer-reviewedIndependent checkStatus
Co-Scientist — AML repurposing, liver-fibrosis targets, AMR mechanism✅ Nature (2026)✅ Wet-lab validation by collaborating labs; AMR hypothesis matched an independent studyStrong, preclinical
ERA — 40 genomics methods beat experts; 14 COVID models beat the CDC ensemble✅ Nature (2026)◐ Public benchmarks; retrospectiveStrong on benchmarks
AlphaEvolve — new algorithms, 23% kernel speed-up, compute recovered in Google's fleet◐ Technical report◐ Mathematical results checkable; deployment internalStrong, narrow
Robin — ripasudil as a dry-AMD candidate✅ Nature (2026)◐ Lab experiments by humans; no clinical trialPreclinical hypothesis
The AI Scientist-v2 — workshop paper passes review✅ System paper in Nature (Mar 2026)◐ 1 of 3 papers above threshold; withdrawn by designWorkshop level
Anthropic AAR — 0.97 PGR vs 0.23 for humans✗ Lab publication✗ Not yetPromising, single task
OpenAI — Navier–Stokes finite-time blow-up with Lean proof✗ Not refereed◐ Lean check exists; mathematicians still checking statement and priorityContested / pending
OpenAI — automated research intern, 3.1 agent-workdays per human workday✗✗ Self-reportedDirectional
AI doing R&D in one paragraph
By September 2026 AI takes part in nearly every stage of research, and in domains with objective feedback — code, benchmarks, formal proofs, assays — it has produced peer-reviewed results: Co-Scientist's validated hypotheses, ERA's expert-beating scientific software, AlphaEvolve's algorithms, Robin's drug candidate. Multi-agent systems now run end-to-end loops, and frontier labs use them for substantial internal R&D. But every system still starts from a human-defined problem and ends at human-run verification; the strongest claims (Navier–Stokes, lab productivity) await independent checking; and full autonomous research organisations — let alone ones that improve their own models — are not demonstrated. As generation becomes abundant, the scarce inputs are research taste, trustworthy validation and the governance of what gets pursued.

Next: How Self-Improving Agentic Systems Actually Work →

⚠️ Notes

Corrections to the source infographic: (1) Anthropic AAR human baseline — the infographic says humans recovered 22% of the gap; Anthropic reports 23% (0.23 PGR over seven days) vs 0.97 for the nine AARs (~800 cumulative hours, about $18,000); (2) OpenAI's math result is overstated — "reported solution to Navier–Stokes equations with formal proof" should read: a claimed proof, with a Lean formalisation, that a smooth 3D Navier–Stokes flow can blow up in finite time; announced 8 Sep 2026, produced by ~10,000 agents over 88 hours (≈2.7M messages, ~130B tokens), not refereed as of mid-September and subject to a priority dispute; it is not a Millennium-Prize award; (3) Robin's result — the infographic cites "KL001 in retinal cells"; we could not find that. Robin's published discovery is ripasudil, a glaucoma drug, as a candidate for dry age-related macular degeneration, with human scientists running the experiments; (4) publication status added — Co-Scientist, ERA, Robin and The AI Scientist were all published in Nature in 2026. Autonomy-ladder shares, the stage map and the funnel model are our illustrative readings.

🔗 References