Text-to-SQL

Two-agent SQL team with read-only Analyst, schema-managing Engineer, and a self-learning loop.

The Dash team is a working text-to-SQL system with a structural separation between the agent that reads data and the agent that builds reusable views.

from agno.team import Team, TeamMode

dash = Team(
    id="dash",
    name="Dash",
    mode=TeamMode.coordinate,
    model=MODEL,
    members=[analyst, engineer],
    db=agent_db,
    learning=dash_learning,
    add_learnings_to_context=True,
    share_member_interactions=True,
)

Two members and one leader, sharing a single learning store.

The split

Agent	Schema access	Tools
Analyst	public (read-only via default_transaction_read_only=on)	SQLTools(read_only=True), introspect_schema, save_validated_query, ReasoningTools
Engineer	public (read), dash (read+write, blocked from public writes via SQLAlchemy event listener)	SQLTools (full), introspect_schema, update_knowledge, ReasoningTools
Leader	None directly	Routes the request, optional SlackTools for posting back

The boundary is enforced at the database engine, not in the prompt. The Analyst’s connection physically cannot write, and the Engineer’s writes physically cannot touch public. These hold even if the model goes off-script.

How a question flows

1. User: “What’s our MRR trend by plan?”
2. Leader retrieves shared learnings (error patterns, gotchas) and routes to Analyst.
3. Analyst retrieves table metadata + matching query patterns from knowledge.
4. Analyst writes SQL grounded in the retrieved patterns.
5. Analyst executes against public (read-only).
6. Analyst interprets results, returns insight.
7. Leader synthesizes and replies.

If the Analyst hits a recurring need (a query the Engineer should turn into a view), it asks the Engineer. The Engineer creates dash.monthly_mrr_by_plan as a view, records it in knowledge, the Analyst prefers it next time.

The self-learning loop

When the Analyst hits a SQL error and finds a fix, it saves the pattern via LearningMachine:

Error: "active subscription detection wrong"
Diagnosis: filtered on status='active', but trials had status='active' too
Fix: filter on ended_at IS NULL instead
Saved as learning: "active means ended_at IS NULL, not status='active'"

Next time the same shape of question hits, the learning gets retrieved before the Analyst writes SQL. Same error doesn’t happen twice.

After a few weeks of use:

• Knowledge holds the validated query library, table descriptions, business rules.
• Learnings holds the error patterns and fixes the team has discovered.
• The Analyst writes correct SQL faster because both are in context.

Why coordinate mode

TeamMode.coordinate means the leader decides who handles each request. Other modes:

Mode	Behavior
coordinate	Leader routes to one specialist per turn
route	Same, but with explicit routing rules
broadcast	Leader sends to all specialists, gathers responses
tasks	Leader breaks the request into tasks, fans out, synthesizes

For text-to-SQL, coordinate is the right shape. One question maps to one specialist (Analyst for queries, Engineer for schema work). The leader picks based on the user’s intent.

What runs where

File	Purpose
agents/dash/team.py	Team definition, leader instructions
agents/dash/agents/analyst.py	Read-only Analyst
agents/dash/agents/engineer.py	Schema-managing Engineer
agents/dash/knowledge/	Loaded at boot via load_knowledge.py
agents/dash/scripts/load_data.py	Generates the synthetic SaaS dataset

See it in action

@Dash what's our current MRR?
@Dash which plan has the highest churn rate?
@Dash show me revenue trends by plan over the last 6 months
@Dash which customers are at risk of churning?
@Dash build me a view for monthly recurring revenue by plan

The first four go to the Analyst. The fifth goes to the Engineer.

For a from-scratch tutorial that builds a Dash-style agent on your own data, see Tutorial: Dash.

Source: agents/dash/

Next

Investment Team →