Agent Steps and Instructions
Iterate the plan conversationally
You'll be able to
- Refine an agent's steps by talking to the builder, not editing a form
- Write a clear per-step instruction that constrains what the step does
Read first
Day 2 Lab
**Status: Generally available — Glean’s Agent Builder and its core features (scheduled triggers, versioning, native actions) are GA. Only the *content triggers* trigger type remains Beta; confirm availability in your tenant.**
- Open yesterday's draft agent and look at its ordered steps.
- In chat, tighten one step's instruction — name the exact channel, time window, or output format.
- Ask the builder to add a guard step: "only continue if there are new posts."
- Reorder the steps if the plan reads out of sequence.
- Read the revised plan back end to end before you let it run.
Working Example: Tightening a Step by Chatting
Per the Fall '25 launch announcement, you iterate the plan in conversation: each step carries an instruction you can sharpen in plain language. Verified against fall25_launch.
Loose step: "Summarize the posts."
You: "Summarize only posts since the last run, max 5 bullets, link each source."
Revised step instruction:
- Scope: posts in #incidents newer than last_run_at
- Limit: at most 5 bullets
- Format: one bullet per incident, each with a source link
- Guard: if zero new posts, stop and send nothing
You: "If there are no new posts, don't DM me at all." -> adds the guard aboveA vague instruction ("summarize the posts") leaves the agent to improvise; a tight one fixes scope, limit, format, and a stop condition. That precision is a MANAGE control — it's how you keep a running agent inside the lines you meant. Remember this builder is announced, not confirmed-GA: verify the step-editing behavior in your tenant before you depend on it.
Check for understanding
What makes a step instruction tight enough to be safe to run — give one property.
Check yourself
1. In Glean's vibe-code agent builder, how do you start building an agent?
2. The vibe-code builder, 100+ native actions, scheduled triggers, looping, and versioning are best described as...
3. Before adding a native WRITE action (e.g., update a Salesforce case) to an agent step, what should you check?
4. An agent runs unattended on a daily schedule. Which guard most directly keeps a bad run from compounding?
5. Why does rolling an agent back to a known-good version count as a governance control?
In 10 seconds
“An agent's plan is a list of steps you can talk into shape. Add a step, reorder, or tighten an instruction by chatting — then read the revised plan back before you let it run.”
The 100+ Native Actions