AgentQ Best Practices

Guidance for getting good answers out of AgentQ without burning through your provider budget. The hard numbers behind everything here live in AgentQ Limits.

Prompt Design

Be Specific About Assets

AgentQ resolves asset references by name. The more specific you are, the faster it finds the right datastore, container, or field, and the fewer tool calls it spends looking.

Instead of...	Use...
"Check my main table"	"Check the `orders` table in the `ecommerce_prod` datastore"
"Fix the data quality issues"	"Create a not-null check on the `customer_id` field in the `orders` container"
"Show me anomalies"	"List active anomalies in the `transactions` container in `sales_db`"

When a name is ambiguous, AgentQ searches for it first and then queries. Giving the full path skips that round trip.

Describe Business Intent, Not Just Operations

AgentQ works best when you say what you want to achieve rather than which rule to build. It translates intent into the rule type and parameters on its own.

Good: "Make sure order totals are always positive and under 1,000,000"
Unnecessarily technical: "Create a between rule on the total_amount field with min 0 and max 1000000"

Open AgentQ From the Asset

Open the floating chat from a datastore, container, field, quality check, or anomaly page and AgentQ picks that asset up as context, shown as a badge above the input. You can then ask "why is this failing?" without naming anything. This is the cheapest way to be specific.

Break Complex Work Into Steps

One message can only go so far before it hits the per-message request ceiling. Sequential messages also let you correct course halfway instead of discarding a long run.

Good: "First, show me quality scores for sales_db." Then, once you have read the answer, "Now create checks for any container scoring below 80."
Risky: "Analyze all 12 containers in sales_db, create quality checks for each, run scans, then send a Slack summary."

Let the Workflow Tools Do Multi-Step Work

AgentQ has five guided workflows that handle multi-step work more reliably than an ad-hoc chain of individual tools. You do not call them by name, you just describe the goal and AgentQ picks one.

Say something like	And AgentQ runs
"Create a check that enforces this rule"	Quality check generation
"Build me a table that joins these two"	Dataset transformation, routed to table, file, or join
"Why did this anomaly happen?"	Anomaly investigation
"How has quality moved this quarter?"	Trend analysis
"What does this score actually mean?"	Quality score interpretation

Managing Cost

AgentQ runs on your own provider account, so every message has a real price. Three habits account for most of the difference.

Start a New Chat for Unrelated Work

A chat carries its history into every later turn, so a long session keeps paying for context it no longer needs. Starting a new chat resets that entirely and is the most reliable fix for a session that has wandered. AgentQ compresses long chats on its own, but compression is damage control, not a substitute for scope.

Narrow What You Ask For

Name the containers you care about rather than asking for all of them.
Ask for counts and filtered lists when you want a number, not the underlying rows.
Do not ask for an entire table. Listings are paged, so a request for 10,000 rows just makes AgentQ walk through them one call at a time.
Ask AgentQ to validate a query before creating a computed asset from it, rather than creating and deleting until it is right.

Stop a Turn That Has Gone Wrong

If the first tool steps show AgentQ heading somewhere you did not intend, click Stop instead of waiting it out. What was produced so far is kept, and the rest of the turn is not paid for.

For the rate limits, token ceilings, and timeouts these habits work around, see AgentQ Limits.

Sending Documents

When your provider accepts attachments, use Attach file rather than pasting a document into the message. The file reaches the model as a document instead of as a wall of text, and the input stays readable.

When it does not, paste the content. Anything you paste over 1,000 characters is captured as a PASTED attachment automatically rather than filling the input box. See Supported AI Providers for which providers take attachments.

Setting AgentQ Up Well

These two choices are made once by a Manager or Admin, and everyone feels them.

Write a Business Context Worth Sending

The Business Context travels with every prompt from every user, so it is the highest-leverage text in the whole setup. Describe your domain, what your team is responsible for, and what your data is for. Keep it to that: it is sent to your AI provider on every single turn, so it must not contain secrets or anything the provider should not process.

Pick the Lowest Level That Still Works

The access level applies to the whole deployment, so raising it to unblock one team exposes source values to everyone. Start from what your team actually needs, confirm which capabilities each level turns on, and raise it deliberately rather than by default. See Access Controls.

Staying Inside the Guardrail

AgentQ declines requests that fall outside data quality, governance, and the Qualytics platform, so a prompt that wanders off topic costs you a round trip. Keep the ask anchored to an asset, a check, an anomaly, or an operation. Short follow-ups such as "yes" or "that one" are read as continuations and pass through on their own. For what the guardrail allows, what it blocks, and how it reads follow-ups, see How AgentQ Works.

Let Long Operations Run

When AgentQ starts an operation such as a profile, scan, or sync, it waits for that operation to finish before moving to whatever depends on it. You do not need to check back and send a follow-up.

That means you can ask for the whole chain in one message:

"Create a computed table joining customers and orders on customer_id, then profile it and create quality checks for the key fields."

AgentQ creates the table, triggers the profile, waits for it to complete, and only then creates checks on the profiled fields.

Review What AgentQ Produces

AgentQ writes for real. A check it creates starts asserting, and a computed asset it creates is a real container. Its answers also vary by provider, model, and the context available at the time.

Read the Answer basis note on an answer before acting on it. It says whether a tool returned the fact directly or the model put the answer together.
Have a person who knows the domain review generated checks and transformations before they gate anything that matters.
Treat a suggested cause, impact, or next step as a lead, not a finding.

Keep the Record

Archiving a chat takes it out of your active list but keeps everything. Deleting one removes it for good, and its turns leave the AgentQ Audit with it. When a chat led to a change in the platform, archive it rather than deleting it, so the reasoning behind the change stays available.