Adopt AI in your business through a practical first project
Choose a first business AI project, prepare data and responsibilities, then compare accepted outcomes before expanding AI adoption across your team.
8 min read
Research teams need to decide what to try, observe what happens and preserve what they learn. AI agents can explore sources, prepare experiments and compare results. Their contributions become useful when conclusions connect to a source or an experiment.
Organize research around a question and a decision criterion. The outcome can be a prototype, a reasoned decision to reject an approach or a clear statement of what remains unknown.
Turn an idea into a research question
“Find a better architecture” leaves better undefined. “Reduce search response time without reducing relevance” supports a measurement, reference dataset and comparison.
Separate known facts, assumptions and open questions. An agent's proposal does not become a project fact when another agent repeats it. Preserve information status through handoffs.
Define exploration boundaries: available time, environment, permitted datasets and decisions the project must preserve. The AI project brief can provide the entry point.
Assign research responsibilities
An explorer finds options and sources. An experimenter turns a hypothesis into a trial. A verifier examines the method and attempts reproduction. A decision role compares results against the mandate. The pilot keeps track of work and prepares the next step.
These responsibilities do not always require five agents. The most useful separation is often between the author of an experiment and whoever independently checks its conclusion. The supervisor workflow guide explains these handoffs.
Scroll to read all columns
| Contribution | Input | Expected output |
|---|---|---|
| Exploration | Question, constraints and known sources | Options, links, dates and unknowns |
| Hypothesis | Candidate and current baseline | Prediction and a result that could refute it |
| Experiment | Protocol and environment | Configuration, measurements and observations |
| Verification | Result and reproduction procedure | Confirmation, discrepancies or inconclusive result |
| Decision | Comparable outcomes and criteria | Reasoned decision and next action |
Define the experiment before using tools
Describe inputs, configuration and measurement. Explain what changes between trials and what stays constant. Changing the model, documents and retrieval method together makes the resulting difference difficult to explain.
Calibrate the instrument against a known case. A script counting the wrong objects or a timer started after an expensive step can return a plausible but incorrect measurement. Include that calibration in the protocol.
Keep failures and incomplete outcomes. A trial that never loaded its data does not measure the quality of an approach. It identifies a protocol or environment blocker to resolve before comparing options.
A team wants to improve answers to project questions. It keeps documents, model and questions constant while changing how passages are selected.
The dataset includes answerable questions, missing information and conflicting documents. Review checks the answer, cited source and handling of absence. Total time includes retrieval and generation. A faster result that is less faithful to its sources is not automatically preferred.
This example specifies a comparison rather than reporting a benchmark performed by Stellary. Execute it on the team's actual sources and configurations before drawing a conclusion.
Explore alternatives without multiplying noise
Give each explorer a different angle: the current method, a technical alternative, operational constraints or earlier work. Request concise, comparable outputs. Several summaries of the same source set do not necessarily represent independent perspectives.
The pilot reconciles options and contradictions, then selects candidates worth testing. Research can end in a focused question for an expert or a decision not to build.
Multi-agent coordination helps track dependencies. An experiment relying on shared data is not independent from a branch changing those data.
Reproduce before optimizing
An improvement should be recoverable with the same inputs and settings. Record versions, uncontrolled variation and attempts. One observation may open an interesting direction without establishing a general improvement.
The optimizer starts after result validity is established. It can reduce cost, latency or complexity while preserving functional criteria. The verifier then checks that an apparent gain did not move a failure into a case missing from the first trial.
For model selection, compare cost per accepted outcome and retries. Local models also require recording hardware, quantization and engine capabilities.
Keep decisions and negative results
A rejected approach remains valuable when the team knows why it failed. Preserve the hypothesis, protocol, outcome, limitation and conditions that would justify another attempt. The next agent can avoid repeating the investigation without reason.
Keep conclusions within scope. “This configuration failed on these documents” is more precise than “this technology does not work.” A project decision is not a universal ranking.
Documentation and agent memory connect this learning to future work. Review a conclusion before promoting it into a lasting project rule.
Run research in Stellary
Create a card for the main question and contributions for the necessary experiments. Link sources and protocol. Define each result's acceptance criteria and the decision expected at the end of the mandate.
A workflow can move a hypothesis from exploration to experiment and independent verification. A condition can request a protocol correction or route the result to a decision. Steps must reflect the actual tools and execution capabilities available.
A project status report separates finished experiments, observations awaiting verification and open decisions. Research also progresses by reducing uncertainty and preventing unnecessary investment. Preserve that value even when a trial ends without a new product.
Build a decision record people can use
The final record should distinguish rejected options, remaining possibilities and the recommended next step. Link every option’s outcome to its protocol. A reader should be able to find the data and understand why the observation matters to the decision.
Separate an environmental limitation from a limitation of the approach. Failing to access a required service does not establish that an algorithm is poor. It may still make the option unsuitable for this project if access is a lasting constraint.
A decision can select the next experiment instead of a final architecture. When two options remain close, identify the test most likely to distinguish them. The pilot can turn that uncertainty into another bounded contribution.
Preserve source quality
An explorer should distinguish primary evidence, commentary and promotional examples. Keep the link and date, then inspect whether the cited passage supports the conclusion. Repeating one claim across several articles does not create independent experiments.
Treat an announced capability as an announcement until tested. Provider results can help choose candidates without measuring your data, constraints or operating cost. Keep those evidence levels separate once your own experiment is complete.
Decide what another team would need to reproduce
Write the minimal reproduction procedure while the experiment is still fresh. Include input preparation, configuration, executed steps, expected output and any known nondeterminism. Reference the exact artifacts instead of vague descriptions of the latest version.
Ask a verifier who did not create the experiment to follow that procedure. Their difficulty can reveal missing setup, an ambiguous measure or a conclusion dependent on undocumented judgment. Correct the protocol before treating the experiment as repeatable.
Stop exploration when it no longer changes the decision
More sources and more agents do not always reduce useful uncertainty. Define the condition under which the current evidence is sufficient for the next decision. A small reversible choice may need a smaller experiment than a major infrastructure commitment.
If research stalls, report the unresolved question and the input needed to continue. That is more useful than producing another synthesis of the same material. Preserve accepted findings so that new information can restart the work from the right point.
Give the first experiment a precise output
For a first task, compare two ways of retrieving a decision from test documentation. Use the same questions, keep the retrieved passages and check whether each answer cites the correct source. The result may be “no improvement demonstrated”. This is a proposed protocol, not a completed trial.
The evaluation worksheet records conditions and observations. Stellary agents can fit into this process; sources and decisions stay in the knowledge base.
