Imbue open-sourced Catalyst to keep AI research agents exploring competing hypotheses

Why it matters

Catalyst turns verification and competing hypotheses into first-class parts of an AI research workflow. If the approach generalizes beyond nanochat, small research groups could run broader computational searches without surrendering control to a closed agent platform, though token and experiment costs remain a practical constraint.

Imbue open-sourced Catalyst on July 20, giving researchers an open-source system that preserves competing hypotheses while agents run experiments, evaluate results and search for better solutions. Imbue resurfaced the work in an August 7 thread that pointed back to the July release instead of announcing a new version.

Imbue on X

Catalyst addresses a problem central to Josh Albrecht and Kanjun Qiu: an AI agent can produce useful work while remaining difficult for its user to inspect, redirect or trust. Qiu studied computer science at MIT, worked at the MIT Media Lab and became Dropbox's chief of staff before founding the Y Combinator-backed recruiting software maker Sourceress. Albrecht was an early Addepar engineer, founded an acquired 3D injection-molding software business and built Sourceress with Qiu. They founded Imbue in 2021 around the idea that AI tools should remain open, modifiable and under their users' control.

That origin matters for Catalyst. The system does not ask researchers to trust a single agent's chain of reasoning. It maintains a population of interpretations, runs candidate experiments, branches into alternative explanations and gives more resources to the approaches that score well. The code is available under the AGPL-3.0 license, so researchers can inspect and modify the machinery orchestrating the agents.

Evolution as an escape from agent tunnel vision

In the launch research, Daniel Mewes (@danielmewes) described a recurring failure mode in linear research agents. Once an agent decides that its main ideas are exhausted, it can repeatedly test small variations instead of reconsidering the premise that led it into the dead end. Imbue calls the resulting behavior "hypothesis collapse."

Catalyst tries to prevent that collapse through "interpretation strands." Each strand contains an agent's working explanation of the evidence and a roadmap of experiments it considers promising. Catalyst samples among those strands, asks agents to propose experiments and executes the proposals judged most likely to advance the research.

Periodically, Catalyst consolidates the observations into updated theories. Some theories branch into deliberately different interpretations. Candidate strands then receive scores based on the performance of the solutions they produce and the novelty of their proposed research paths. Separate review agents look for reward hacking, specification violations and evidence that a proposed result fails its verifier.

The approach uses familiar language-model agents as components. Catalyst's contribution is the search and evaluation process around them. Its repository supports Claude Code, Gemini CLI, Antigravity CLI and Codex CLI, with users supplying their own model access and paying the providers directly.

The nanochat result comes with a narrow denominator

Imbue demonstrated Catalyst on nanochat, the small transformer training project used by Andrej Karpathy's AutoResearch. Karpathy, who joined Anthropic earlier in 2026, designed AutoResearch around a coding agent that repeatedly changes a training setup and retains modifications that improve the measured result.

The test gave each experiment five minutes on one H100 GPU. Agents could modify nanochat's architecture, training recipe and hyperparameters, while a verifier measured validation performance in bits per byte. Imbue says Catalyst's evolution workflow reached a val_bpb score of 0.9361 after 340 experiments and had not plateaued when the run stopped.

Imbue describes the improvement as three times farther than a regular AutoResearch agent. That figure comes from one nanochat optimization setup, rather than a benchmark spanning different models or research problems. Imbue also compared Catalyst with work from Recursive Superintelligence, though the comparison required estimating Recursive's experiment count from wall-clock time. Recursive changed from an H100 to a B200 during its run, and Imbue used only the H100 portion. Imbue acknowledges that Recursive may have reached comparable scores more efficiently.

Those caveats narrow the claim to something useful and defensible: evolutionary search kept finding improvements after the tested linear agents stalled. The result does not establish that Catalyst will outperform other research systems across scientific domains.

Open code leaves a substantial compute bill

Catalyst is best suited to questions whose answers can be tested by code. Its repository recommends narrow optimization targets, reproducible computational phenomena and explicit verification scripts. Imbue says the current implementation is a poor fit for physical experiments, underspecified goals, general engineering projects and problems beyond the underlying model's capabilities.

The Catalyst repository sets a default limit of 30 minutes for a single experiment, although users can change it, and says evolution workflows frequently invoke more than 100 subagents. Its cost table estimates that a Develop Theory evolution run costs about $200 with Gemini 3.5 Flash or $1,000 with Claude Opus 4.8. It estimates a Solve Verifiable Goal evolution run at about $500 with Claude Opus 4.8. Imbue also says review and scoring consume roughly 65% of the tokens in a typical Develop Theory workflow. These API estimates exclude the compute used to run experiments.

That allocation shows where Qiu and Albrecht are placing their bet. Better research agents will require spending much of their budget challenging, ranking and revising ideas instead of generating a single polished answer. Verification becomes part of the core workflow instead of a final check performed after the agent finishes.

Catalyst's more ambitious theory workflow follows the same principle. In a companion experiment, Imbue used populations of candidate explanations, adversarial reviewers and falsification attempts to study a neural-network training phenomenon. Imbue warns that hallucination, data falsification, missing citations and reward hacking remain unresolved. Human researchers still have to monitor the work and verify its conclusions.

Catalyst follows Imbue's 2026 reset

Catalyst also reflects an organizational shift inside Imbue. In a January 30 account of the reset, Albrecht wrote that he and Qiu had become disillusioned after pushing Imbue through an intense sprint to ship Sculptor, its parallel coding workspace. He said they wanted Imbue to move toward smaller projects built in the open, with individual project owners given broad control over what they shipped.

Catalyst makes that plan concrete. It is an inspectable research project with named contributors, reproducible experiments and an invitation for outside researchers to contribute templates. The Catalyst repository names Daniel Mewes, Catherine Kim and Evan Ryan Gunter as contributors.

Imbue has the financing to pursue that model without turning every release immediately into a paid product. In 2023, Imbue announced a $200 million Series B at a valuation above $1 billion, with participation from Astera Institute, Nvidia, Cruise co-founder Kyle Vogt, Notion co-founder Simon Last and other investors. Imbue later added $12 million from Amazon's Alexa Fund and Eric Schmidt.

Catalyst is a technically narrow release relative to that capital base. Its importance rests in the operating model Qiu and Albrecht are testing: open research software that lets humans specify the goal, inspect the search process and challenge the agent's conclusions. The nanochat result gives that thesis an early data point. Broader research use will determine whether competing hypotheses can help agents produce discoveries that survive scrutiny outside a controlled optimization task.