Agents

An agent is an AI colleague you hire and brief yourself. It has a name, a job, a model behind it, a set of things it’s allowed to touch, and standing instructions in your words. Unlike Adlyse AI — which is one general assistant shared by everyone — an agent is yours, narrow on purpose, and can be pointed at a schedule or dropped into a workflow.

In one line

Build an agent when a recurring job deserves an owner with a standing brief. Keep it narrow — three tools and one job beats twenty tools and a vague mandate.

Before you build one#

Run the job by hand in Adlyse AI first. If you can get a good answer out of a single conversation, you now know two things: the job is possible, and you have a prompt worth pasting into the agent’s instructions. If you can’t get a good answer by hand, an agent won’t rescue it — fix the brief first.

See when to use which for the fuller comparison against workflows and the built-in assistant.

Creating an agent#

Go to AI Studio → Agents → New agent. The editor is split: configuration on the left, a private test conversation on the right.

Give them an identity#

Three fields, and the agent isn’t created until all three are filled:

Name — what you’ll call them, e.g. Maya.
Role — their job title, e.g. Creative strategist.
Description — one line on what they do, e.g. Drafts weekly performance reports from Meta & Google Ads.

This isn’t decoration. The role and description are how you and your teammates pick the right colleague out of a list six months from now — and how a parent agent decides whether to hand this one work.

Select Create agent. The rest of the editor, and the test conversation, unlock.

Pick a brain#

The Brain is the model behind the agent. Heavier models reason better on open-ended analysis and cost more per run; lighter ones are fine for well-specified, repetitive jobs. If you’re unsure, leave the default and change it once you’ve seen real runs.

Add integrations#

Integrations are the specific data sources and actions this agent can reach: your ad platforms, a knowledge base, an MCP server or a single MCP tool, a workflow it can call, a custom API.

An agent with no ad-platform integration cannot read your performance data, no matter what its instructions say. This is the most common reason a new agent answers “I don’t have access to that.”

Choose sub-agents#

Sub agents are other agents this one can hand work to. Use them when a job genuinely splits — a report writer that delegates the Meta pull to a Meta specialist — rather than stuffing every capability into one agent.

Write how they work#

How they work is the standing brief: their job, their style, what a great answer looks like, and the rules they must never break. Write it the way you’d brief a new hire on day one.

Be specific about the things you’d otherwise have to repeat every conversation:

plaintext

You are our ecommerce performance specialist.
Target ROAS is 3.2. Never propose changes to brand-term campaigns.
When you flag a campaign, always include 30-day spend and the trend.
Report in plain language — the reader is a founder, not a media buyer.

Turn on only the capabilities it needs#

At the bottom, Capabilities are the tool groups — memory, folder access, charts, documents, creative studio, suggestions, data tables. They’re off by default and you should keep it that way until a real run proves one is missing.

See tool groups for exactly what each one grants and where it works.

Test it in the panel beside you#

The right-hand panel is a private trial conversation with your draft. Ask it the real question — the one you’d ask on a Monday morning — and watch what it reaches for. Adjust the brief, ask again.

Running an agent#

Once it’s saved you can:

Chat with it from the Agents page, for a one-off.
Schedule it — see Schedules.
Call it from a workflow, as an Agent node inside a larger pipeline.

Scheduled agents propose, they don't apply

An agent running unattended cannot change your ad platforms directly — every real platform change needs an approval preview, and on a background run there’s nobody there to approve it. Give the agent the Suggestions capability and it will write proposal cards into your decision queue instead, with the numbers and reasoning attached. Brief it to propose, not to fix.

Seeing who does what#

The Agents page has two views. Cards is the default roster. Hierarchy draws the delegation map — who hands work to whom — which is the fastest way to spot an agent that’s quietly become a bottleneck, or two agents that overlap and should be one.

Adlyse AI itself appears in the roster alongside your own agents, so you can see how the safest configuration in the product is put together.

Common flows#

Promote a good conversation. You asked Adlyse AI something that worked; paste that prompt into a new agent’s brief so it happens weekly.
Build a read-only reviewer. A compliance agent with folder access and no ad-platform write ability — it can read everything and change nothing.
Split an overloaded agent. If the hierarchy view shows one agent doing four jobs, give two of them to sub-agents with narrower tools.

Tool groups explained

What each capability toggle grants, and why some only work in chat.

Put it on a clock

Run an agent daily, weekly, or hourly without being the trigger.

Orchestrate several

When the order of steps matters more than the judgment in them.

Teach it a routine

Save a repeatable task once so it runs the same way every time.

For developers

Create and manage agents programmatically with create_assistant and update_assistant, and run one with run_assistant.