AI Agents are Hygraph's intelligent automation assistants designed to handle repetitive content workflows, such as translation, summarization, and SEO optimization. Powered by integrated large language models (LLMs), agents understand your project's schema and context, applying automation to specific fields or entries based on the configuration you define.
You configure an agent once and trigger it manually on selected entries, or let it run automatically when an entry moves through a workflow step.
Every agent run is reviewable. When an agent completes, the entry enters a read-only review state. You can approve or revert each field the agent changed before the entry returns to normal editing.
Hygraph uses the latest Claude models to power AI tasks such as SEO analysis, content summarization, and translation. These models are selected for their quality, accuracy, and efficiency.
We are continuously testing new models to improve performance and reliability. Model versions may be updated over time as newer releases become available. Updates are applied automatically to maintain optimal results.
#Availability and token usage
AI Agents is available on the Enterprise plan only. Each project receives a set number of AI tokens per billing period. You can view your current usage and limits at any time on the Billing page.
#Agent types
#SEO Expert
Reviews all entry fields, including nested components, and posts a structured SEO report as a comment on the entry. The report covers content optimization, heading structure, and metadata gaps. Relation fields are excluded. Guidelines can be assigned to shape the agent's output. See AI Guidelines.
Examples
- Automatically review every blog post before it moves to the publish step and receive a structured SEO report as an entry comment.
- Identify missing meta descriptions or weak heading structures across all entry fields.
- Open a draft entry and trigger the SEO Expert from the sidebar to get a report before moving it to the review step.
#Translator agent
Translates content fields into one or more target locales, preserving tone, structure, and meaning. Custom instructions and Glossary guidelines let you control formality, handle untranslated terms, and adapt output to specific audiences. Relation fields are not translated. See AI Guidelines.
Examples
- Translate all localizable fields into French, German, and Spanish as soon as a content entry is approved in the source locale.
- Apply custom instructions such as
Keep product names untranslated to ensure consistency across all target languages.
- Select all blog posts in a specific locale and run the Translator agent to generate translations in one pass before a product launch.
- Translate a single entry on demand when a localization request comes in outside of the normal workflow.
#Content summarizer
Generates concise summaries from one or more source fields and writes the output to a target field on the same entry. Tone, length, and focus can be guided through custom instructions. Source and target fields are configured when launching the agent, not during setup. Guidelines can be assigned to shape the agent's output. See AI Guidelines.
Examples
- Generate a 2–3 sentence excerpt from a long-form article body and write it to the
summary field automatically.
- Produce concise overviews of product descriptions for use in listing pages or meta descriptions.
- Run the Content summarizer across an entire content type to populate a
summary field that was previously empty.
#How to trigger agents
Each agent is configured for one triggering mode — agent runs or workflows. These are mutually exclusive.
#Agent runs
The agent is available for manual triggering from the content table or the content entry form. Select one or more entries and trigger the agent from the action bar, or open a single entry and trigger it from the Agents panel in the sidebar.
#Workflows
The agent runs automatically when an entry moves through a configured workflow step. No manual action is required once the workflow is set up.
#Safety guardrails
Agents only write to the fields and models you explicitly configure. They cannot delete entries, unpublish content, or modify your project structure.
The review step means no agent output reaches your published content without an editor sign-off.
#Review agent output
When an agent run completes, the entry displays a Ready for review status and the content entry form becomes read-only. The agent's changes are already written to the entry. The review step is where you decide what to keep.
The review presents a side-by-side diff of the original and agent output, field by field. You can approve or revert individual fields. Reverting all fields returns the entry to its exact pre-agent state.
The entry stays read-only until the review is complete. This applies both while the agent is running and while the completed run is awaiting review.
#Best practices
Workflow agents
- Position agents strategically in your workflow. Run agents during pre-publish or review steps so they operate on finalized content and produce the most accurate results.
- For multilingual projects, pair the SEO Expert with a Translator agent. You can run SEO analysis on the source locale first, then translate the content so localized content benefits from the same SEO feedback.
- Run the Content summarizer to generate a short excerpt, then pair it with the SEO Expert to validate the metadata before publishing.
- Always include a human review step after an agent step in workflows handling sensitive or regulated content. Agents produce output to review, not output to publish automatically.
- Monitor AI token usage on the Billing page to stay within your plan limits. Avoid configuring agents on high-volume workflows without first estimating token consumption.
Agent runs - bulk updates in the content table
- Agent runs are limited to 50 entries at a time. Before running an agent on a large batch, test on 3–5 representative entries first. Review the output and confirm the agent behaves as expected before scaling up.
- Check the Ready for review status in the content table to track outstanding reviews after a bulk run. Entries in read-only state block editors from making updates. Clear the review queue before triggering follow-up runs.
- Scope agents to specific models rather than all models. An agent available across all models appears in every content table, which creates noise for editors working on models it was never intended for.
Agent runs - single entry updates in the content entry form
- Use single-entry runs from the content entry form to test a new agent configuration before using it for bulk updates. The side-by-side review gives you a clear picture of what the agent changes before you roll it out at scale.
- If an agent run produces unexpected output on a specific entry, check whether that entry has unusual field values, such as empty fields, very short content, or fields in an unexpected locale, before adjusting the agent configuration.
General
- Review agent configurations periodically to ensure custom instructions and assigned guidelines remain accurate. Brand voice, legal requirements, and audience context change over time.
#What's next
Ready to configure your first agent? See the AI Agents setup guide.