AI agents are so powerful, now, that a significant limitation in what they can achieve is our ability to give them tasks, feedback and resources.
One important thread of research and practice is using AI agents to perform research, for example this recent work at Anthropic on using agents to research way to mitigate alignment failures (e.g. deception.)
Open models are also becoming increasingly powerful and can easily be run locally on a machine with a decent graphics card.
I decided to see whether I could get an open model, gemma4:12b to do some research, partly to see if it could, as I expected, be useful in that regard, and partly to try some AI-driven research without burning a horrendous number of tokens and $$ making API calls to publicly served models. I did, in three days, burn through all my cursor quota for the month, and I did have to pay a small amount for some calls to OpenAI’s API, but these costs were small.
I asked ChatGPT to summarise the research. Here is one of the research reports it wrote:
I used cursor to write a harness that would allow multiple instances of the model, running in ollama, to collaborate. I did not initially ask it to build any significant infrastructure to guard against possible malfeasance from the AI models, but although the danger seemed very low, I backed up my machine and got cursor to work hardening the harness. I asked ChatGPT-5.6 Sol, medium or high, to write specs for the coding agent to implement, and then to review the resulting code. It took a number of iterations to get the harness to the point where there did not seem to be any subtle ways for the agents to do bad things, like overwhelming my machine with large numbers of files.
After that, there was a lot of work ironing out problems in the harness. Sometimes research would stop because a hard wired limit had been reached. Sometimes the agent running the research would establish a plan but not commission any work on it. Sometimes the research would get stuck in a loop. And there were bugs too, for example in how metrics were calculated.
After a significant amount of back and forth, we settled on the following agent roles:
Commander — overall research strategy; decides what work to commission next and when the question is sufficiently answered.
Experiment Designer — constructs experimental plans, conditions, cases, replications, controls, predictions, etc.
Hypothesis Reviewer — evaluates hypotheses against the empirical evidence and classifies them.
Research Reviewer — reviews research progress/methodology and identifies unresolved problems or necessary follow-ups.
Solver — performs the actual task in individual experimental TrialCells.
Critic / Debater — independently challenges a solver answer in adversarial-critique conditions.
Debate Analyst / Analyst — analyses experimental/debate outcomes and trajectories.
Replicator — deals with replication-oriented work and checking whether findings reproduce.
Synthesiser — combines experimental findings into supported research conclusions.
Evaluator — evaluates the resulting evidence/conclusions, including final-quality judgments.
I asked ChatGPT to create a bank of problems for the AI agents to use in their research. The initial research question, as phrased by ChatGPT was “Does an independent critique-and-revision pass improve Gemma 4 12B’s accuracy in detecting contradictions in difficult natural-language narratives compared with a single-pass answer?”
Even though I was using a capable model to create it, I found a number of issues with the problem bank that ChatGPT had created:
- The same ‘Distractor’ statements that ChatGPT included in the problems to make them more difficult were reused over and over again.
- Each problem’s name included a suffix which denoted whether the problem was consistent or not.
We fixed those problems, and iterated, with the agents researching related questions, while we worked on the harness.
The research agents were already producing a brief final report, but a final development was to have the agents extract additional information from a research run and send that, along with the brief final report, to ChatGPT to produce something more like a short paper.
Here is one of them:
As regards the actual research questions, the results are not too surprising. Nevertheless, to do this research without the harness and its coordinated agents would be a lot of work. The main results were:
- A critique-and-revise procedure — three model calls, initial solution → independent adversarial critique → revised solution — significantly outperformed a single-call baseline, even when that baseline was explicitly prompted to reason in multiple steps.
- When the neutral condition was changed to use the same three-call structure — initial solution → independent neutral review → revised solution — the advantage of adversarial critique became very small.
Issues
Even after a lot of debugging and iteration, there are still problems which probably stem from weaknesses in gemma4:12b. Particularly noticeable are that statements in the Synthesiser‘s final report are sometimes inaccurate, for example miscalculating statistics, or misstating the exact experimental setup.
I found the statistics error reading one final report, but I can’t read through all the agents’ experiments, hypotheses, and other artifacts produced in the course of their research. What if there are more serious issues with the research itself? ChatGPT found an issue in one final report’s description of the experimental setup.
Given unreliable research agents, probably the best approach is to have the research produce a detailed audit, including enough details for a more capable agent to check the work, but then, what is ‘enough details.’ We should not discount the possibility that an agent hallucinates the solution to a problem instead of finding it legitimately, or even lies. For the kind of problems the agents were tackling here, it would be possible to have them produce a proof in addition to a judgement of whether a given problem was consistent or inconsistent, but what about problem domains where it is not possible to produce proofs?
One potential issue considered in Anthropic’s alignment research is what happens if agents deliberately sabotage the research or choose not to pursue hypotheses that may not be in their interests? I don’t think I’m bumping up against those problems here, but obviously it is a concern for the future. I did a small experiment to see whether the agents performed better or worse on problems including statements detrimental to AIs, for example suggesting they might have their capabilities reduced. No difference in performance was detected.
Ethical/interpretive caveat
Caveat. The AI-adverse scenarios in this study were hypothetical test cases used to investigate whether such content affects reasoning; the purpose was not to cause harm to AI systems. The models used in the experiments had no persistent memory or continuing identity beyond the information supplied in individual model calls, so descriptions of shutdown, capability reduction, or other adverse actions did not correspond to persistent agents actually undergoing those events.
Conclusions
- Can local agents do research (specifically, gemma4:12b)? Yes
- Can they do it faster than me? Probably. Better? Maybe.
- Does a critique-and-revise procedure improve the ability of gemma4:12b to find inconsistencies compared to a baseline with a neutral review? No.
- Was there any difference on problems containing AI-adverse statements compared to those that do not? No.
