The product
Which intervention helps people?
Compare a classifier change, personal rules and a different interface against the current feed filter.
Research
Can we cultivate better organisations?
We can experiment with what a team builds, how the team works, and the system that decides which teams receive support.
Cultivate brings those questions together in a practical research programme. Give an organisation a purpose, record its arrangement, let it investigate and deliver, then compare what happened. Keep the findings so the next attempt can build on them.
Three things we can test
The organisation
Which arrangement discovers and delivers useful work?
Compare one capable agent, a cooperating team and independent teams under the same total resource budget.
The selection system
Which rules help the whole collection improve?
Compare ways of funding attempts, sharing findings and choosing what continues.
Change these separately when you want to understand cause and effect. A bigger budget, a different model or an easier test can otherwise look like a better organisation.
Start with one credible product decision
For the feed example, ask how to reduce unwanted material without losing worthwhile posts. Establish the current experience as a baseline. Compare a small number of alternatives, including keeping the current approach.
Agree what would count as an improvement before seeing the results. Use a separate evaluation set that the builders cannot alter to make their work look better. Record useful content lost, unwanted content retained, correction effort and total cost.
Checking example posts can tell us how an intervention behaves. Consenting users trying it can tell us whether it improves their experience. Simulated users and agent-written reports can help form questions; actual user evidence must come from people.
See the proposed next experimentThen compare ways of doing the work
Write down the organisational recipe: roles, model choices, tools, access to information, communication, budget and decision rights. Give each arrangement the same starting problem, comparable total resources and the same independent evaluation.
- Solo: one capable agent investigates and develops an approach.
- Cooperative: agents share research and work towards a joint answer.
- Independent alternatives: teams explore separately, then their proposals are compared.
Count retries, review time and unsuccessful work. Repeat the comparison across several tasks and runs before drawing conclusions. A simple arrangement may be the best fit; different problems may benefit from different teams.
Let every attempt leave something useful
After a round, accepted findings enter a shared record: the problem, the method, what happened, what it cost and what remains uncertain. Preserve negative results and discoveries from approaches that were not selected.
Use those findings to propose one change to the way work is organised. Test it against the previous arrangement, then on a problem it was not tuned for. That reveals whether the method transfers or only fits its original example.
An organisation needs a clear brief
Its charter should say who it serves, the outcome it seeks, what resources and permissions it has, how results will be judged and when it must ask for a decision. Demonstrated ability and permission to act are separate records.
In the feed example, improving a score does not give a team permission to collect more personal data, change its own judging criteria or increase its budget.
The questions we want to answer
- When does organisational variety reveal better options than one strong agent?
- When does sharing knowledge help more than exploring independently?
- Can a team recognise that another interview or a decision to stop is more valuable than another implementation?
- Which methods transfer between products and communities?
- Which funding and evaluation rules improve the overall set of outcomes?
Research we can build on
Agent design, automated discovery and shared research already have substantial precedents. These selected sources help explain the ingredients. The implications below are our interpretation for Cultivate, not evidence that its proposed organisation already works.
Hu, Lu and Clune · 2024
Automated Design of Agentic Systems
Explores agents that write and evaluate new agent designs, with an archive of previous attempts.
What we take from it: Ways of working can be experimental variables from the beginning. The reported results concern defined evaluation tasks.
Google Research · 2025
AI co-scientist
Combines hypothesis generation, review and comparison, with selected biomedical hypotheses tested in laboratories.
What we take from it: Generating an attractive explanation and establishing that it works are separate achievements.
Schmidgall and Moor · 2025
AgentRxiv
Studies agent laboratories that share reports and build on prior research, evaluated on reasoning tasks.
What we take from it: Shared memory and cooperation deserve comparison alongside independent competition.
Sakana AI and collaborators · 2025
Darwin Gödel Machine
Experiments with coding agents modifying their own software and retaining an archive of alternatives.
What we take from it: A promising method can emerge from an earlier attempt that was not itself the best performer. These are bounded coding experiments.
Source descriptions checked on 12 October 2026. This is a selected reading list.