Research together, build on evidence

Use CoAutoResearch to advance work you can understand, question and revise. The agent can carry out a sequence of research steps; you bring context, decide what matters and remain responsible for interpreting the result.

Give recursive self-improvement a clear target

Define what a better method or research decision would mean before comparing alternatives. Keep evaluation rules, resource limits and approval conditions explicit. Let the agent use recorded outcomes and review feedback to revise the next step, including abandoning an unsupported hypothesis.

Capture reusable lessons where they can inform later work; a repeated log of completed commands adds little value. Discuss questionable assumptions in a parallel chat and send guidance when you want it applied. The improvement loop should accumulate useful knowledge, not merely additional trials.

Start with a useful brief

An initial question is enough to begin a conversation. A stronger brief explains what is known, what remains uncertain and which constraints matter. Add a proposal, prior experiments or notes when available. Say which conclusions would change your next decision, rather than requiring a positive result.

For example: “Compare these two methods under a fixed compute budget. Report uncertainty and failure cases. Do not use the final test set to choose settings.” A target venue helps shape reporting, but does not establish novelty or supply missing evidence.

When an action needs a later decision, name that action and its prerequisite. Review the prepared brief before Start to confirm the condition was retained. Starting autonomous work does not waive a separate approval you reserved.

Bring existing work without losing its meaning

Upload research material or link an accessible folder through the composer. Uploads are copied; linked material stays in place. Include enough context to distinguish raw data, derived results, abandoned experiments and current code. Do not include credentials or unrelated private files.

Research-bearing material in resources/ongoing_work/ triggers a conversion trial before ordinary research. It inventories existing work, records resource locations and treats earlier quantitative results as tentative until checked. An imported number is not automatically a verified new finding. Preserve source files and explain how earlier results were produced.

Use discussion to improve the next step

Open New chat while autoresearch is working. Ask a specific question: Is the comparison fair? Would the finding survive a different split? Which alternative explanation remains? Discussions can inspect the project without directly changing its research files.

When an answer suggests a useful change, choose Add to research draft. Review the wording and send it to the research session. Make the intended scope explicit: “Before the next evaluation, add class-level error analysis” is easier to apply than “make the study better.” Check the following plan or result to see how the instruction was handled. Adding a suggestion to a draft does not send it to the research session.

Keep each trial focused

A trial should resolve a concrete uncertainty or remove a blocker. Define a bounded next move with an observable outcome: acquire data, compare methods, repair an experiment, investigate a failure, or organize verified findings. Prefer a small useful run before spending more compute.

For statistical work, establish evaluation rules in advance. Fit preprocessing only on training data, preserve an untouched test set where appropriate, record seeds and account for uncertainty. Do not repeatedly change splits until a result looks attractive. A careful negative comparison can be more useful than an unreliable improvement.

Read live updates for the current action, observations and next step. Review technical details when investigating a specific problem. If a run stalls, check whether it is waiting for a provider, executing work or asking for a decision; a failed page refresh is not itself evidence of a failed experiment.

Resolve execution setup before experiments

Ask the agent to verify its project-local interpreter and dependencies before using evaluation data. If a copied Python environment cannot start, inspect the error and try an installed interpreter with the required version. Do not treat a failed installation as an experimental result or silently inherit system packages when isolation was requested.

Codex web search and command networking are separate settings. If dependency downloads fail, inspect the actual error and host connectivity first. When the cause is the workspace command network restriction, you can enable downloads for this project in Settings → Agent → Advanced config with sandbox_workspace_write.network_access=true, save, and resume the stopped run. This enables network access for commands, not just the browser search tool. See the official Codex network documentation.

Make resource searches useful

Ask which missing source could change the method, comparison or interpretation. Resource Scout can inspect relevant papers, dataset documentation and code; its reports are available in Resources → Research source reports. Check what was actually read, how it affects the next step and which gaps remain. Finding a citation is not the same as verifying the claim it is used to support.

Reuse adequate source checks. A local experiment may still need better method or evaluation grounding, while a purely local repair may need no new search. Judge source coverage and relevance rather than bibliography length. New literature does not justify repeating a used final test set or presenting post-result exploration as an analysis specified in advance.

Give guidance at the right boundary

Use Pause after current turn when you want to review progress before automatic continuation. Wait for Paused before assuming work has stopped. Resume with specific guidance when the current direction remains useful.

Use Stop current run when an active execution must be interrupted; its work may be incomplete. Reserve Restart autoresearch for returning to the original launch direction, with later work archived. A follow-up message usually preserves more useful context than editing and resending an earlier message.

Read findings with their limits

Distinguish a proposed idea, a measured observation and a reviewed result recorded in the project. Inspect data provenance, comparison conditions, uncertainty, failed attempts and unresolved explanations. Internal review checks the research record; it does not substitute for independent replication or peer review.

Candidate routes are directions still available to explore. Rejected or replaced routes belong to history and help explain how the current direction was chosen. Avoid treating the number of trials or review files as a measure of scientific progress. Ask what the latest evidence actually allows you to claim.

Move from manuscript to paper in the product

Use Manuscript to read the evolving research story and inspect its evidence. The underlying manuscript/PAPER_PLAN.md describes the writing plan; manuscript/BLUEPRINT.md connects content to supporting material. Missing analysis belongs in the research process, not in invented prose.

Once reviewed results are recorded and project agents are idle, choose Generate paper. The product’s internal agent uses scientific writing, visualization, citation and venue-template skills to produce a draft and PDF from the saved evidence. It does not run new experiments during writing.

Choose a general research report when the work is an exploratory comparison. For a conference target, specify the venue and year. A template or a long PDF cannot compensate for missing evidence. Review notes should distinguish issues that better writing can resolve from gaps requiring further research.

Check every figure against its source data. Numeric figures should preserve uncertainty, units and comparison conditions. Conceptual illustrations must not be presented as experimental evidence. Then inspect citations, author details, limitations, ethics statements and the actual venue requirements before sharing. See paper generation for tools, cancellation and recovery.

Model connection stalls

During Codex research, a model request with no progress for 20 minutes is stopped so the dashboard does not wait indefinitely. Diagnostic retries do not extend this deadline. An explicitly running tool is allowed to finish under its own resource limits. This is a runtime interruption, not evidence against the research idea. Review Activity and resume the unfinished phase when the connection is available; CoAutoResearch does not automatically spend another Trial to retry this timeout.