← AIAP Tool GuidesAll programsHomeSearch
AIAP Tool Guides·Lyzr·56:31

Lyzr Agent Studio Deep Dive: a No-Code Manager Agent with Specialist Sub-Agents, Knowledge Bases, Testing and Deployment

Lyzr deep-dive instructor (unnamed) Consultant/instructor building agents for clients; solo recorded screen-share, not Lyzr staff

The short version

  1. Multi-agent design: a manager talks to the customer and delegates to narrow sub-agents, each with its own knowledge - like company teams (0:00-0:03).
  2. Build for a fictional "Awesome Shopper" store: two knowledge bases (catalogue, return policy), a Return Specialist and a Shopping Assistant (role, goal, instructions, memory), then an Orchestrator with GPT-5 Mini (0:03-0:34).
  3. Test in the inference panel (a 35-day sunglasses return correctly refused under a 30-day policy; an outfit request came back incomplete), then deploy as an app page, an API call or an MCP server (0:37-0:56).

At a glance, three clicks deep

Skim here first: the closed row is the glance, open is the study card with the key points and timestamps, and the ↓ link drops to that concept's full write-up below.

01Manager agent plus specialist sub-agentsOne flexible manager routes each request to the sub-agent whose role matches;›

One flexible manager routes each request to the sub-agent whose role matches; sub-agents answer only from their own knowledge and hand back what is out of scope.

Concept (0:00-0:03)

Sub-agents (0:12-0:24)

Manager (0:24-0:34)

↓ Full write-up of this concept

02Knowledge bases, agent fields, testing and deploymentCreate a knowledge base per topic, define each agent by name, description, role, goal and instructions, tes…›

Create a knowledge base per topic, define each agent by name, description, role, goal and instructions, test with the trace panel, then launch.

Knowledge bases (0:03-0:10)

Testing (0:37-0:45)

Deployment (0:46-0:49)

↓ Full write-up of this concept

The concepts in full

01

Manager agent plus specialist sub-agents

Keep each role "extremely simple and extremely direct" - the manager reads roles to decide who gets the job.

Sub-agents keep the manager toggle off and little or no memory; the manager gets the list of sub-agents with a one-line purpose each, instructions saying when to call which (without that it "will not know who to delegate"), no knowledge base and ten messages of memory. Overlapping roles - a shipping agent next to a returns agent - confuse routing unless clearly separated. Model choice: "start with the mini models" and move up only if answers are weak.

Why it matters

The structure behind most production support agents.

02

Knowledge bases, agent fields, testing and deployment

how-to

If behaviour is off, "the first thing which you touch is always the prompt."

Knowledge base: folder > Create > name plus a description that says when to use it > upload PDF/DOC/TXT (up to five files at recording). Return Specialist instructions: "if order exists show status, if not say it doesn't exist, ask to recheck or raise ticket." Shopping Assistant goal: "narrows down options, gives best deal, make sure the user feels confident". The Improve Prompt button lightly rewrites. Testing shows delegation in an activity log. Deploy: Launch agent (private marketplace page), Agent API (curl), or the Lyzr MCP server so agents become tools in any MCP client.

Do it in this order
Why it matters

A repeatable no-code path from documents to a deployed agent.

Tools referenced

ToolCoverageMomentContext
Lyzrdemonstrated
GPT-5 Minidemonstrated
n8nmentioned
Claude Codementioned
Weaviatementioned
Qdrantmentioned
Amazon Bedrockmentioned
Hugging Facementioned

Session materials

Archived locally on V: — click to open. Companion pages link to the LMS.

Action items

    Resources mentioned

    Resources
    • docLesson page on the platform

    Extraction notes

    This page was built from an auto-generated transcript, which garbles product and people's names. Those were corrected silently in everything above and logged here for transparency. The warnings flag claims that were true on the recording day but change fast.

    Transcript corrections applied

    The transcript saysThe trainer actually means
    Lyser / Liza / LyzaLyzr
    studio.lyza.aistudio.lyzr.ai
    VV8Weaviate
    QuadrantQdrant
    any 10n8n
    cloud codeClaude Code

    True on recording day — verify before relying