An AI agent reads your site's content through the admin's own data tools, forms a judgement, and records what it found. A proofreading agent reads a blog post as it is published and proposes corrections. An SEO agent audits a page's title and meta description and suggests better ones. A sales agent summarises yesterday's orders each morning and posts the summary to Slack.
Agents do not act on their own. They run when a hook or a scheduled task invokes them, they read only the data their allowed tools expose, and, unless you deliberately configure otherwise, everything they propose waits for a person to approve it.
How this differs from Core dna MCP
Core dna also lets you drive the platform from an external AI assistant, covered in the Core dna MCP via Claude / ChatGPT / Claude Code pages in the Integration Docs. That is the opposite direction: you sit in an AI tool and it reaches into Core dna on your behalf.
AI Agents run the other way. Core dna calls the AI provider itself, on a trigger you configure, with no one at the keyboard. The two features share the same underlying tool layer, but nothing you set up for one applies to the other.
The four things to keep separate
Library agent: a ready-made agent shipped with the platform, such as Page Meta Audit or Sales Report. Library agents are the same for every client. You browse them under Agents ▸ Library.
Install: your configured copy of a library agent. The install holds the settings you chose: its parameters, which model tier it uses, which AI provider it calls, how much autonomy it has, and who to notify when it needs help. One library agent can be installed once per client.
Invocation: one recorded run of an install. Every run is logged with its transcript, token usage, cost, decision summary and outcome, under Agents ▸ Invocations.
Recommendation: one proposal produced by a run. Most agents record several per run. Recommendations queue under Agents ▸ Recommendations, where you accept or reject them.
The shape is always the same: a library agent is installed once, each install runs many times as invocations, and each invocation records zero or more recommendations.
How much autonomy an agent has
Each install has an action mode, chosen when you install it:
- Read-only (recommendations only): the agent proposes; nothing is ever written. The safest mode and the right default.
- Accept to action: proposals wait in the Recommendations inbox until a person accepts them, at which point the change is applied.
- Auto-apply with veto window: proposals apply automatically once a delay you set has passed, unless someone rejects them first. You also set a confidence threshold, below which a proposal still waits for a human.
Not every agent offers every mode. An agent that only reports, such as Sales Report or Order Anomaly Triage, is read-only and cannot be given write access.
What an agent can read and change
Agents read your data through a fixed, per-agent list of tools, and can write only to the fields the agent declares. Both lists are set by the library agent and cannot be widened, not by configuring an install, and not by forking one to a custom prompt.
Three capabilities are built in to every agent:
- record_recommendation: propose something for review.
- emit_result: return a structured result that later hook actions can read.
- ask_human: pause the run and ask a person a question. See Escalations.
Model tiers
You do not pick a specific model. Each install has a model tier, Small, Medium or Large, and the tier is resolved to a real model for whichever provider the install uses:
| Tier | Anthropic | OpenAI | |
|---|---|---|---|
| Small | claude-haiku-4-5 | gpt-5.6-luna | gemini-3.1-flash-lite |
| Medium | claude-sonnet-5 | gpt-5.6-terra | gemini-3.7-flash |
| Large | claude-opus-5 | gpt-5.6-sol | gemini-3.1-pro-preview |
Every library agent recommends a tier. Leave it on Recommended unless you have a reason not to. Larger tiers reason better and cost more per run.
What you need before agents will run
- An AI provider key for your client, saved under Agents ▸ Providers & budgets. Agents use your own provider account; Core dna does not resell model access.
- An installed agent, configured from the library.
- A hook or scheduled task that invokes it. Agents never run on their own and cannot be triggered by hand from the agent screens.
Optionally, a connected Slack workspace if you want escalations and reports delivered to a channel.
Where to go next
- Providers & budgets: add a provider key and set spend caps. Do this first.
- The agent library: see what is available and install one.
- Installing an agent: the configuration choices in detail.
- Reviewing recommendations: the day-to-day review loop.
- Invocations: the run log, for when something did not work.