Build Glean agents end-to-end

Glean is an enterprise AI platform built on knowledge retrieval from 100+ connected sources — Confluence, Jira, Salesforce, Google Drive, Slack, GitHub, and more. Agents are defined as structured JSON workflows with typed steps: search, LLM inference, API calls, form triggers, and schedules.

An agent can run in Workflow Mode (deterministic, schema-validated steps) or Auto Mode (the model decides tool use), and can be triggered by chat, an input form, a schedule, or a webhook.

AESOP runs its build phases against a Glean-specific fitness check, then generates the artifacts Glean needs. The pipeline is:

• Fitness check — score the PRD against Glean’s rubric: are the data sources indexed, is the trigger supported, does the output channel fit.
• Workflow design — lay out the typed steps: search → synthesize → respond or act.
• Write instructions — emit the workflow JSON and any extension pieces.

The extension pieces are where Glean gets richer than a single prompt. AESOP can generate Skills (reusable SKILL.md packages), Custom Actions (REST calls from an OpenAPI spec), Action Packs (pre-built integrations), and Sub-agents (published agents called as workflow steps with isolated memory).

Once built, the workflow is published to Glean. AESOP then evaluates it with its Glean connector, which talks to the deployed agent over POST /rest/api/v1/chat using the agent’s agent_id and holds a chatId across turns for continuous multi-step conversations.