Skip to main content

​
Why prompt management matters

An agent’s quality depends heavily on how well its prompt is written. As you learn more about your data and how the agent behaves, you’ll want to iterate on the prompt — adding context, correcting misunderstandings, or refining instructions. Interact tracks every prompt change as a versioned history, so you can safely iterate without losing previous working configurations.

​
Prompt versioning

Every time you save an agent, Interact creates a new prompt version. Versions are numbered sequentially and timestamped.

​
Viewing version history

  1. Open an agent and click Edit agent
  2. Go to the Prompt history tab
  3. You’ll see a list of all saved versions with the date and a preview of what changed

​
Restoring a previous version

If an update makes the agent perform worse, you can roll back:
  1. In Prompt history, find the version you want to restore
  2. Click Restore — this loads that version’s prompt into the editor
  3. Click Save — a new version is created with the restored content
Restoring creates a new version, it doesn’t overwrite history. You can always see the full history of changes.

​
Learnings

Learnings are a record of what your agents have worked out about your data, and of the feedback users have given on their answers. You don’t create them by hand — they accumulate on their own as agents run and as people rate responses.

​
Agents save their own learnings

When the Save learning tool is enabled for a subagent, the agent decides for itself when something is worth writing down and saves it mid-conversation. It’s told to reserve this for things with real substance, not minor details. Typical triggers:
  • Schema structure it had to work out — union tables, sparse rows, what a column actually contains
  • Conventions it inferred — naming patterns, units, how tables join
  • A correction from you that is reusable — “revenue is in cents”, “that table joins on client_id”
  • Tool and formatting insights — how a particular chart or report needed to be shaped
You’ll see the save happen as a step in the tool panel while the agent works.
If an agent isn’t recording anything, check that Save learning is enabled for the relevant subagent in the agent builder.

​
Feedback on a response becomes a learning

This is the part users drive. Any response can be rated, and the rating is recorded alongside the agent’s own learnings, tagged User feedback:
  • Thumbs up — recorded as a preference, marking the answer as a good example
  • Thumbs down — opens What went wrong?, where you pick a category and can add details
The details box asks “What should the agent have done differently?” — answering it properly is what makes the feedback useful to whoever reviews it later.

​
Reviewing learnings

All learnings for the organisation are listed under Agents → Settings → Learnings, newest first. Filter by All, User feedback, or Agent learnings, and expand any row to read the full text. Each entry links back to the agent, and to the chat or job run it came from, so you can see the exact conversation that produced it. This view is read-only — learnings aren’t edited or deleted from here.
Learnings are a record for review, not a live feedback loop. They are not injected into an agent’s context, so saving one — or leaving negative feedback — does not by itself change how the agent behaves next time.To make a fact stick, put it where the agent actually reads it: the agent’s context or instructions for agent-specific facts, or the organisation context for facts that apply everywhere. Use the Learnings list to find out what to write there.

​
Where to put a durable fact

For facts that apply everywhere — standard metric definitions, company naming conventions, market codes — use the organisation context instead of duplicating them across agents.

​
Best practices

The orchestrator prompt should say in one sentence what the agent is for. Everything else follows from that. Vague prompts (“help with marketing data”) lead to inconsistent behaviour.
Don’t try to document every table upfront. Start chatting, and when the agent queries the wrong table or misinterprets a column, correct it in the conversation and then add that correction to the agent’s context or a table note. Build context incrementally through real usage.
Dates are one of the most common sources of agent errors. If your reporting week starts on a Monday, say so. If “last month” means the full calendar month (not rolling 30 days), say so explicitly in the prompt.
Context and table notes are best for specific facts: column meanings, ID formats, units, edge cases. Prompt text is best for behavioural rules: always show results by campaign, always present ROAS to 2 decimal places, always flag anomalies.
Rather than guessing what your prompt is missing, read what agents have had to work out for themselves and what users have flagged. Recurring entries are the strongest signal for what belongs in the prompt or context permanently.