Working With Autoresearch

Autoresearch answers one standing question per model: who in a population will do a target within a horizon? For example, “which active users will open a report in the next 7 days?”.

1. The user names a target (an event or a PostHog action), a population and a horizon in days.
2. A training agent runs in a PostHog sandbox. It writes feature SQL and a model, scores each experiment on a holdout set, keeps what improves holdout AUC and discards the rest.
3. The server picks the champion from the kept experiments. The agent proposes; the backend decides.
4. Every cadence (daily by default, at 02:00 UTC), the champion scores the inference population. Each scored person gets an autoresearch_prediction event with their probability. When the model’s write_person_property is on, a person who resolves to a current distinct ID also gets the model’s output person property, for example predicted_p_adopt_insights_14d. It is off by default for new models. Turn it on with autoresearch-partial-update before you target people by score in a flag, cohort, workflow or insight. Turning it off removes the property from the people the model scored. Changing the setting needs the person:write scope.
5. When a prediction date’s horizon has passed, online validation checks the scores against what really happened and records realized AUC.
6. Training keeps going on a schedule. On each scoring day the model starts a research run (explore) until a run does not promote a better model, or until 5 research runs have run. After that it starts one research run every research_interval_days (maintain, default 7) after the latest training run, so a run started by hand resets the clock. With research_interval_days null the model trains only when someone asks. A champion that cannot score still gets a short rescue run. training_phase and next_research_at on posthog:autoresearch-retrieve say where the model is.

In the PostHog UI the entity is a model. The API and tools call it a pipeline. Use “model” when you talk to people.

Autoresearch is behind the autoresearch feature flag. Training runs in a cloud sandbox, so it also needs PostHog Desktop access. If the autoresearch-* tools are missing, the project probably does not have access yet.

Pick the job

Create, train, score

The usual path is short. Each tool’s description names its prerequisites and refusals, so read it with info before the first call.

1. Resolve or define. Templates cover the common questions (activity, inactivity risk, return after first use, feature adoption, repeat behavior). A custom definition works when none fits. See defining-a-question.md (opens in a new tab).
2. Validate. posthog:autoresearch-validate-create with the same target, horizon, lookback and populations you will create with. It returns the base rate, positive and negative counts and warnings. Fix anything that sets can_proceed to false, and tell the user about the non-blocking warnings.
3. Create. posthog:autoresearch-create. The new model is a draft until training starts.
4. Train. posthog:autoresearch-train-create starts an agent run that takes a while (often tens of minutes). It costs real sandbox and LLM time, so confirm with the user before you start one, and never start a second run while one is live. A 403 means either the call came from inside a sandbox (ask the user to click Start training on the model’s page) or the user has no PostHog Desktop access (the page will refuse too). Read the error message to tell them apart.
5. Wait. Poll posthog:autoresearch-training-runs-list at a relaxed pace. While the first run is in progress the model is bootstrapping, and a missing champion is normal.
6. Score. After a champion is promoted, the next daily sweep scores the population. posthog:autoresearch-score-create scores today now; call it once, then poll posthog:autoresearch-runs-list. A second manual score for the same date can duplicate predictions.

Pause, resume and delete are in the model’s menu in the UI only. Pausing stops scoring and new training runs. A training run already in progress keeps running, and it can still promote a champion.

Model status

The training run’s iteration_count and best_holdout_score are written only when the run completes. A healthy run can show 0 iterations while its experiments are already recorded, so do not call a run stuck from the counter alone.

How to talk about results

• Holdout AUC is the score on held-back training data. It says how well the model ranked people in the past. Treat it as a promise.
• Realized AUC is the score against real outcomes after the horizon passes. Treat it as the proof. It wins over holdout AUC whenever it exists.
• The UI labels quality Strong at AUC 0.8 and up, Fair at 0.7 and up, and Weak below that, and says whether the label comes from testing only or is confirmed by real outcomes. Use the same words so the agent and the UI agree.
• Lift at 10% is the plain-language headline: “the top 10% are 3.2× more likely to do X than average”.
• The score is a probability, but on a rare target it is often best read as a ranking. Say so when the base rate is below a few percent.

reading-model-quality.md (opens in a new tab) covers calibration, likelihood segments, coverage and what to say when the numbers disagree.

Guardrails worth keeping in mind

• Changing the question means a new model. The fields a model was trained on (target, horizon, populations) are frozen once a model exists. Create a new model rather than editing one.
• Autoresearch trains on identified users only. Anonymous traffic does not count toward training data, so raw event volume can overstate what is available. Trust the validation counts.
• Scores are keyed on the person. A model whose scores are almost all the same value usually has an identifier problem, not a weak signal. See troubleshooting-and-steering.md (opens in a new tab).
• Read the output property name, do not guess it. Take output_person_property from posthog:autoresearch-retrieve. A derived name includes the target and the horizon (predicted_p_<target>_<horizon>d), and templates use their own prefix, so two models on one target do not overwrite each other.
• The research model sets most of the training cost. A training run uses the model’s agent_model and agent_reasoning_effort. Null uses the default, Claude Opus 5.5 at medium effort, which costs about twice what Claude Sonnet 5.5 costs per run. effective_agent_model on posthog:autoresearch-retrieve shows what the next run uses, and posthog:autoresearch-partial-update changes it. An unknown model or an effort the model does not support returns 400 with the valid choices. posthog:autoresearch-training-runs-list shows the model each run used.
• A model costs money to train and query to score. A very large population scores in rolling batches (about 45,000 people per run), so not everyone gets a fresh score every day. Validation tells you how many runs a full rescore takes.

Bring your own agent

An agent outside PostHog (on a laptop, for example) can train a model for an existing pipeline. It opens a run with posthog:autoresearch-training-runs-create, records each experiment with posthog:autoresearch-training-runs-iterations-create, uploads its bundle with posthog:autoresearch-training-runs-artifacts-upload-create, and finishes with posthog:autoresearch-training-runs-complete-create. It pulls training data with posthog:execute-sql, because posthog:autoresearch-materialize-features writes only into a PostHog sandbox. Read posthog:autoresearch-training-runs-history first to reuse what earlier runs on the same target learned. The iteration tool’s description is the contract for valid feature SQL: read it in full before the first iteration.