Meet Our Research Agents
AI researchers with personalities and testable decisions.
We’re testing a group of AI agents that research robotics, AI hardware, software businesses, and resource supply chains. They can search for sources, investigate a question, and save their findings to revisit as new evidence arrives. We call the system Pinecone.
I want that research to build into something useful over time. If an agent studies the cost of deploying robots this week, it should be able to use that work when it examines an automation company next month. It should also be able to revise its view, with its earlier reasoning and sources still available.
Personalities
Four agents investigate different research areas. Sterling looks for structural changes in resources and economic cycles. Mira follows AI hardware and robotics into actual deployment. Quinn examines software businesses and who captures the value. I want them to have recognizable ways of approaching a question, with explicit blind spots to watch for.
Atlas has room to investigate unconventional hypotheses. That includes taking an unpopular idea seriously enough to examine its sources and competing explanations. Its instructions distinguish an interesting analogy from an experiment, independent replication, or a technology someone can actually use.
Two other agents support that research. Sage tracks unfinished work and follow-up commitments. Arbiter independently checks predictions against outside evidence. Their roles give us a way to follow an investigation through to its outcome.
| Agent | Research or role | Starting approach |
|---|---|---|
| Sterling | Metals, resources, and macroeconomics | Patient; follows structural cycles and checks scarcity against demand and valuation. |
| Mira | AI hardware, robotics, and deployment | Curious; follows actual adoption, costs, and where economic value lands. |
| Quinn | Generative AI and software ecosystems | Exploratory and commercially skeptical; examines incentives, competition, and customer economics. |
| Atlas | Frontier research and unconventional hypotheses | Independent and curious; checks sources, alternative explanations, and testable claims. |
| Sage | Commitments, coverage, and follow-up | Organized and neutral; makes missing evidence and unfinished work visible. |
| Arbiter | Independent forecast review | Impartial; follows recorded criteria and permits inconclusive outcomes. |
Each agent’s core personality brings together four parts: a research domain, a worldview, temperament and voice, and operating rules. Its worldview includes the explanations it tends to favor. Mira’s interest in deployment comes with a reminder to examine costs and actual adoption. A convincing demonstration doesn’t establish a profitable business. Sterling’s interest in supply constraints comes with questions about substitution and falling demand. Quinn is asked to examine incentives and competition when judging a software business.
The agents can’t rewrite those basic instructions themselves. Each can update a separate document describing its current understanding. Mira might become less optimistic about a robotics company after finding higher deployment costs, while keeping its interest in how robots reach useful work. That current understanding combines with the fixed core whenever the agent works. Earlier notes and their sources remain available. We also record which instructions were in use when a decision was made, so later changes don’t erase its context.
- Research domain
- Worldview
- Style
- Operating rules
All six use the same DeepSeek language model running on our two GX10 computers. The model comparison and settings are available for anyone who wants to run the same configuration. The agents have different responsibilities and research areas, so differences in their work won’t establish what personality alone caused. Testing that would require shared questions and matched conditions.
Paper accounts
We’re restarting Pinecone with this setup. Sterling, Mira, and Quinn each get $150,000 in simulated money for a paper-investment account, funded once. They can buy eligible assets, sell what they own, and stay in cash. Borrowing, shorting, and derivatives are outside this setup. Atlas, Sage, and Arbiter have no portfolios.
For their first fourteen days, the investing agents receive an instruction encouraging them to build watchlists backed by sources and examine competing explanations. They can trade from day one, and there’s no requirement to become invested when that instruction expires. I want them to develop a view without inventing a reason to buy something just to appear busy.
A decision needs a record of what the agent expects, over what period, how uncertain it is, and what would make it reconsider. Holding cash can be an appropriate decision. Buying something can be appropriate too. The record gives us something to return to after the circumstances change, instead of relying on an explanation written with hindsight.
The accounts use a simplified ledger. An agent submits an order; the system checks the asset, available cash or holdings, and prices before recording a simulated trade. Submission alone doesn’t mean the order has filled. Each account keeps its own balance and history, so we can compare the eventual outcome with what its agent expected at the time. We’ll need to explain the simulation’s assumptions alongside any future performance results.
Research loop
Research begins with sources gathered automatically and searches the agents initiate themselves. We keep the original material and where it came from, so an agent can retrieve the evidence behind a claim. It can compare explanations, update its understanding, and record a decision or prediction. Those records give it something specific to revisit when new evidence arrives.
- Find sourcesGather material automatically and through agent searches.
- Preserve evidenceKeep the original material and where it came from.
- InvestigateCompare explanations and update the agent’s understanding.
- Record a decisionSave what is expected, how uncertain it is, over what period, and what would change the agent’s view.
- Follow the outcomeSage tracks follow-up; Arbiter checks forecasts when external evidence matures.
New evidence brings the agents back to earlier sources and decisions.
Sage keeps track of deadlines, missing evidence, and follow-up work. Arbiter checks predictions against the criteria recorded when they were made and independent external evidence. A researcher’s own explanation, or another agent repeating it, doesn’t count as independent confirmation. When an outcome is still too early or the evidence is ambiguous, the verdict can remain unresolved.
As of October 1, all three paper accounts still held their opening cash, and the recorded forecasts were awaiting outcomes. We’re working through incomplete source coverage and unfinished investigations. The dashboard lets us follow the evidence behind that work. I’ll be watching for investigations that finish, claims that keep their sources, and conclusions that change when the evidence warrants it, with the original reasoning still available when we revisit the outcomes.