Agent best practices: getting the most out of the system

1. Write goals that produce great plans

The single biggest lever is the goal. You don't need to specify steps — the orchestrator plans those — but naming the outcome, scope and thresholds gets a sharper plan and fewer clarifying questions.

• State the outcome: “pause”, “rebalance budgets”, “build a remarketing audience”.
• Scope it: which accounts, campaign-name patterns, and a date range.
• Give thresholds: “ROAS below 1.5×”, “cost > $20 and 0 conversions”.
• One goal per task — split unrelated jobs so each plan stays focused.
• Let it ask: when the agent auto-clarifies, answer rather than guessing for it.

2. Set guardrails in account memory first

Before you run writes, put your operating rules in per-account memory so every run respects them automatically — you won't have to repeat them in each goal.

• Targets: CPA / ROAS / CTR goals the agent optimizes toward.
• Protected campaigns: patterns it must never pause (e.g. “brand*”).
• Excluded geos and custom rules (e.g. pause if ROAS < 1.5).
• Capture your methodology once; it's applied on every future run.

3. Start read-only, then enable writes

Reads are free and safe — begin with audits and analysis to build trust in the agent's reasoning. Then enable writes, which pause at an approval gate so nothing ships without your yes. For changes across many accounts, always take the dry-run blast-radius preview first.

4. Use approval gates deliberately

• Approve in-app for day-to-day work; set an approver email for stakeholders who don't live in the app.
• Reject a single step to skip it without aborting the whole run.
• Use auto-approve / autonomous mode only for tasks and schedules you fully trust — and only with account-memory guardrails in place.
• Review the audit log to see exactly what changed, when and why.

5. Tune cost with per-agent model choice

Each agent can run on a different model. Keep specialist agents on a cheap ($) model — that's the default and it's reliable for executing the plan — and reserve a stronger model for the orchestrator, where plan quality matters most. If credits run low, drop specialists to a cheaper model to stretch the balance.

6. Make what works repeatable

• Turn a proven run into a Skill (or let the agent build one) so it follows your method next time.
• Pin the Skills a task should use in the composer to steer the plan.
• Schedule recurring jobs on a cron — each fire re-evaluates live data, so it never re-proposes yesterday's actions.

7. Trust the right numbers

Platform-reported conversions can mislead. Join Google Ads cost with your own BigQuery / CRM revenue so pause and budget decisions run off true ROAS — the numbers you actually believe.

8. Scale across accounts safely

• Select multiple accounts in the context picker to fan one goal across an MCC.
• Always preview cross-account writes with a dry-run before applying.
• Lean on per-account memory so each account keeps its own targets and protected campaigns.

Frequently asked questions

How do I get the best results from the agent?

Give a clear goal (outcome + scope + thresholds), set your guardrails in account memory first, start read-only to build trust, then enable approval-gated writes. Keep specialists on cheap models and turn proven runs into Skills and Schedules.

Is it safe to turn on auto-approve?

Use auto-approve / autonomous mode only for tasks and schedules you fully trust, and only with account-memory guardrails (targets, protected campaigns, rules) in place. Otherwise keep the approval gate on every write.

How do I reduce credit cost?

Keep specialist agents on a cheap ($) model and reserve a stronger model for the orchestrator, scope tighter date ranges and fewer accounts, and pin a Skill so the agent follows a shorter proven plan.

How do I make the agent follow my process?

Put your rules and targets in account memory, and capture your method as a Skill (or let the agent build one). Both are applied automatically on every future run.