Putting AI inside a CMS creates a problem: it moves faster than you can build trust for it. This puts your brand, translation, and workflows at risk because for most new tools, you cannot know whether everything will be done as it should be.

Most CMS platforms opted to resolve this problem by moving very slowly: one entry at a time, inline accept-or-dismiss, no audit trail worth speaking of. This defeats the purpose of AI. You want speed, not a mindless clickthrough process.

Others started moving too fast and left a trail of changes that needed human editors to go back and fix them, eroding their confidence in the system and doubling the time spent resolving issues.

This update to our AI agents is the answer to this problem. With today’s release, you can stay fast with your AI and preserve trust and reliability by looping in humans at the critical moments.

Here’s a video walkthrough to learn about the release, quickly:

Watch

#The three agents

We’ve already shipped these three agents, but with this update they are faster and more reliable. They can all be applied at scale, without editing each content entry one at a time. Each agent has a distinct mode of AI-assisted content work:

• The Translator: Translates entries to languages of your choice using your configured guidelines. Creates a version snapshot before touching anything; rollback is always one click away.
• The Summarizer: Takes specified input fields and writes to an output field (e.g. title + body → excerpt). Tone and guidelines are configurable at runtime.
• The SEO Expert: Read-only. Analyzes a page, generates a score, and posts a comment with recommendations. Never modifies content directly.

#How it works: trigger, review, decide at scale

The interaction model is simple:

1. Select up to 50 entries from the content table.
2. Choose your agent.

The agent runs in the background while you work elsewhere. When it's done, the entry gets held in a pending state, visible in the content table and waiting for a human decision.

The review process itself is a version comparison, where you see how every field compares before and after. You can accept all of it, reject all of it, or accept some fields and revert others. Nothing goes live until you say so.

Throughout the entire process, the AI’s work remains versionable, with differences identified, and the possibility of rollback preserved.

#The principles underneath

There are a few things that are true for every agent, every run:

• Agents are no longer confined to workflows. They can be triggered from the content table for bulk operations, or from the content form for a single entry.
• All changes are draft-only; nothing gets published without explicit approval from an editor.
• Every field has full review and revert capability, every time.
• AI Guidelines, including brand voice, glossaries, legal constraints, are set once at the project level.

#A little more on guidelines

One of the most significant changes in the release is where brand context lives. Previously, every prompt had to carry the full weight of your brand rules. Now, guidelines are a first-class concept at the project level.

You can configure up to four categories: Glossaries for translation consistency, Brand Voice & Tone, Audience & Context, and Legal & Compliance. Any agent can inherit up to three of them per run. It means your AI stays within guardrails even when the person triggering it doesn't remember to specify them.

#Enabling large content operations

These changes will benefit large organizations the most. AI-powered content operations for teams are no longer just viable, but easy:

• Content managers can edit hundreds of entries per day.
• Launching hundreds of pages into new markets can take minutes instead of months.
• SEO evaluation and suggestions move from months of agency work to minutes of a single internal process.

#Built to scale responsibly

We’ve been deliberate about how we built this tool, ensuring we take a responsible approach to working at scale. You can run agents across 50 entries at a time, which means using AI at scale, but also, your token costs don’t balloon into unpredictable billing costs.

Beyond cost, we've also decided to acknowledge what AI isn't yet good at: catching its own mistakes, knowing when brand rules have been updated, or understanding that a field currently open in another tab should be left alone. That's why human review remains a key part of every agent run.

If you’re curious about the previous iteration of Hygraph’s agents, you can read about them in the earlier release notes.

Want to see them in action, yourself? Explore our documentation to try them out, or contact your Customer Success Manager to discuss your use case.