Research perspective / Multi-agent systems

Research is a team effort. AI should be too.

Tropistic · · 10 min read

We believe AI research will move forward through small, coordinated teams of agents. Not because more agents are inherently smarter, but because good research needs different questions explored, different assumptions tested, and evidence that survives scrutiny.

Research is not one uninterrupted thought

A useful research project moves between several kinds of work. Someone has to understand the literature, formulate a hypothesis, build an experiment, investigate a surprising result, and decide whether the evidence is strong enough to change direction. These activities inform one another, but they do not always need to happen in one conversation.

That is the opportunity we see in multi-agent systems. By an agent, we mean a model operating in a loop: choosing an action, using a tool, observing the result, and deciding what to do next. A multi-agent system gives several such loops different responsibilities and a way to exchange results. It is not simply the same question sent to several models.

Separate contexts can preserve focus. A literature worker needs papers and source passages. An experiment worker needs code, configuration, and logs. A reviewer needs the claim and the evidence required to test it. Giving every worker the entire transcript can make the system more expensive without making the relevant information easier to find.

Anthropic’s account of its research system describes a lead agent delegating to parallel workers, with particular benefits for breadth-first investigations. It also reports substantial token costs and difficulties with tasks that require tightly shared context. [1] We take both observations seriously. The right unit of parallelism is an independent question, not an arbitrary piece of a prompt.

A small team, with explicit boundaries

We favor a coordinator-and-workers pattern as a starting point. A person sets the question, constraints, and success criteria. A coordinator proposes a bounded plan. Workers explore separable tasks. Their outputs return to a review step before the system presents a conclusion or requests another round.

The coordinator should own the task graph, not every detail of every investigation. It needs to know which tasks depend on others, what evidence is missing, and whether more work is worth its cost. Workers should have narrow responsibilities and enough freedom to choose useful tools within those boundaries.

01 / From question to reviewed result

01 · Direction Human sets the goal Question · constraints · success criteria
02 · Coordination Plan the investigation Dependencies · permissions · budgets

03 · Independent tasks, running in parallel

Literature Find the evidence Sources + assumptions
Experiment Test the baseline Code + configuration
Challenge Probe alternatives Confounders + gaps
04 · Verification Inspect the shared evidence Reproduce · compare · resolve conflicts
05 · Judgment Return to the researcher Accept · redirect · stop

Follow the work from a bounded question to a reviewed result.

Independent tasks can proceed together. Dependent experiments must wait for their prerequisites. Every branch returns artifacts for review.

AutoGen provides a useful foundation for thinking about this division. Its design combines agents with different capabilities, tools, and human participation through explicit interaction patterns. [2] We do not think one conversation topology is universally best. A fixed workflow is often preferable when the steps are known. Dynamic delegation is valuable when the investigation itself reveals what needs to happen next.

A worker brief should answer five questions: What are you trying to establish? What is outside your scope? Which tools may you use? How much time or compute may you spend? What evidence must you return? Clear answers reduce duplicated searches, incompatible assumptions, and tasks that continue after they have stopped being useful.

A worked example: test the method, not the story

Imagine a small team asking whether a new training method improves a baseline under a fixed compute budget. This is a hypothetical workflow, not a result from Tropistic. The first step is to specify what “improves” means: the target metric, datasets, allowed budget, and conditions that would make the comparison invalid.

  1. Map the evidence. A literature worker identifies the original method, its assumptions, and relevant baselines. It returns source passages and unresolved questions, not just a list of links.
  2. Establish a reference. An experiment worker reproduces the baseline in an isolated environment. It records the code revision, dependencies, configuration, seeds, and actual resource use.
  3. Challenge the comparison. A separate worker looks for confounders: different data preprocessing, unequal tuning effort, evaluation leakage, or a metric that hides important failures.
  4. Run the dependent work. Only after the baseline is usable should the proposed method be compared against it. Multiple independent configurations may run in parallel, but their inputs and budgets must remain comparable.

This last distinction matters. A diagram can make every task look parallel; the underlying experiment may not be. If the baseline is broken, launching a large sweep only produces more ambiguous results. The coordinator should wait on genuine dependencies rather than optimizing for the appearance of activity.

Suppose the headline result improves, but the reviewer finds that the new method used more training data. The correct next action is not to write a more persuasive summary. It is to flag the mismatch, revise the experiment, and ask whether the remaining budget justifies rerunning it. A negative or inconclusive result is a valid outcome.

The shared artifact matters more than the chat

Conversation is useful for coordination, but it is a poor substitute for a research record. We believe each consequential claim should point to an inspectable artifact: a source passage, a versioned script, a dataset reference, an experiment configuration, or a run log. The artifact should survive summarization and remain available to the next worker.

A useful handoff separates observation from interpretation. “This run produced this metric under these settings” is an observation. “The method generalizes better” is an interpretation that needs further evidence. Keeping those statements distinct makes it easier to challenge a conclusion without losing the underlying work.

02 / The evidence loop

  1. 01 · Claim A testable proposition State conditions and uncertainty.
  2. 02 · Artifact An inspectable record Source passage, code, or run log.
  3. 03 · Check An independent test Reproduce, compare, or challenge.
  4. 04 · Decision A justified next step Accept, revise, or stop.

Move from a claim to evidence, a check, and a decision.

The record survives even if the conclusion changes. A failed check is a reason to investigate, not a reason to hide the result.

We would keep a compact shared record containing the current question, task status, artifact references, important decisions, and open uncertainties. Workers can retain detailed local context, while the coordinator receives a short summary with links back to the evidence. This is a design preference, not a claim that one memory architecture solves all coordination problems.

Agreement is not an independent check

Several agents using similar models, sources, and prompts can make correlated errors. A second agent agreeing with the first does not turn an unsupported statement into a finding. Useful review reaches outside the conversation: opening the citation, rerunning the command, examining a held-out evaluation, or checking whether the stated goal was actually met.

The MAST paper groups failures into system design issues, inter-agent misalignment, and task verification. Its analysis includes missing checks, incorrect checks, and premature termination. [3] One practical lesson is that adding a verifier role is not enough. The verifier needs explicit criteria, access to the right artifacts, and permission to return “not established”.

Autonomy needs a control surface

A research system should be able to do useful work without requiring a person to supervise every tool call. That does not mean it should have unlimited permissions. We would separate low-risk exploration from consequential actions and make the boundary visible to the researcher.

  • Bound access. Give workers only the tools, files, and credentials required for their task. Run unfamiliar code in isolated environments. Treat instructions found in papers, web pages, or tool output as untrusted data.
  • Bound expenditure. Set limits for tokens, tool calls, compute, retries, and elapsed time. A worker should report a blocker when a limit is reached, not quietly invent a substitute result.
  • Gate consequences. Require explicit approval for publishing, changing shared infrastructure, sending external messages, or exceeding an agreed budget. People should be able to pause or stop work.

Reliability also means surviving ordinary failure. A tool times out. A machine disappears. A worker returns a partial result. Anthropic describes the importance of durable execution, checkpoints, and tracing in long-running research. [1] We believe completed work should be recoverable without replaying the entire investigation.

That requires care around side effects. Retrying a source lookup is different from retrying a compute allocation or an external write. Record whether an action completed before repeating it. Preserve successful artifacts, retry only the failed step where possible, and expose incomplete work in the final report.

Measure the system, not its confidence

A convincing demonstration is not enough. We would compare the team against a strong single-agent baseline using the same task set, access to evidence, and clearly reported resource budgets. If the team spends more, the evaluation should show what that extra spending buys. Parallelism can reduce elapsed time while increasing total compute.

Evidence quality
Are the important claims supported by the cited material? Can a reviewer distinguish findings from assumptions?
Experimental validity
Can someone rerun the work? Are baselines, budgets, and evaluation conditions comparable?
Coverage and calibration
Did the team investigate relevant alternatives? Does it acknowledge uncertainty and missing evidence?
Cost and recoverability
How much time, compute, and human intervention did the result require? What happens when one component fails?

We would also remove components one at a time. Does the reviewer catch errors that would otherwise survive? Do parallel workers discover distinct evidence, or repeat each other? Does shared memory improve recovery? These ablations reveal whether the architecture earns its complexity.

For a simple lookup, a single tool call may be enough. For a short, tightly coupled task, one agent may be easier to evaluate and control. Anthropic’s advice to begin with the simplest effective system is a useful constraint on our own enthusiasm. [4] A multi-agent design should be a measured choice, not a default setting.

The direction we believe in

We believe research agents will be most useful when they resemble a well-run collaboration: focused work, explicit handoffs, independent checks, and people who can change the direction. The goal is not the largest possible team. It is a system that helps a small team ask better questions and follow more of them through to trustworthy evidence.

That is the form of leverage we want to build toward at Tropistic. More room to explore, without giving up the ability to understand what happened, question a result, or decide that a different problem deserves our attention.

This article describes our design perspective, informed by the work below. It does not report original experiments or Tropistic performance results. The diagrams are conceptual, not a depiction of a released product.

Sources & further reading

  1. How we built our multi-agent research system Anthropic, 2025. Architecture, evaluation, and cost trade-offs.
  2. AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation Wu et al., 2023. Sections 2.1 and 2.2 on agents and interaction patterns.
  3. Why Do Multi-Agent LLM Systems Fail? Cemri et al., 2025, v3. Sections 4 and 5 on coordination and verification failures.
  4. Building effective agents Anthropic, 2024. Start simple and add complexity only when it helps.