Information Extraction Agent - Update Agent failed

[Nadav]

Hi everyone,

I’m running into a persistent issue when trying to save an Information Extraction Agent in Databricks.

Every time I click Save and update, I get:

Update Agent failed. Failed to save Agent

This happens regardless of what I change — schema, guidelines, instructions, or even when I try saving without making any edits.

 What I’ve already checked

• The dataset is properly loaded (Unlabeled dataset).

• The selected column is a STRING column with valid text.

• Verified that:

  • No NULL values

  • No empty rows

  • The agent successfully loads sample examples

• Tried creating a completely new agent → same error.

• Tried different browsers.

• No background optimization jobs running.

• JSON schema is valid according to the built-in validator.

• Shortened instructions/guidelines to avoid length limits.

---

My questions

1. Are there known issues with saving Information Extraction Agents?

2. Does the dataset have to be a materialized table instead of a view?

3. Is there a way to see detailed logs for this failure?

4. Are there undocumented validation rules that may block saving (length limits, formatting, etc.)?

Any help or insights would be greatly appreciated. Thanks!

[Louis_Frolio · Accepted solution]

Hello @Nadav 

Thanks for the detailed notes—this looks like the common “Update Agent failed” pattern we’ve seen with Information Extraction agents, and we can narrow it down systematically. Also, the link you shared points to a user profile and I couldn’t open the actual thread; please share the specific community post URL if you want me to reference it directly.

What’s known about “Update Agent failed”

• This feature is in Beta, and we’ve had intermittent backend issues where the UI shows a generic “Update Agent failed” even for transient rate limit or model-serving hiccups. Retrying later sometimes succeeds without any changes.

• Multiple internal tickets and Slack threads document agents failing to save/update due to transient outages and/or rate limiting against the underlying foundation model—again surfacing only as the generic “Update Agent failed.”

• Schema-related issues can also trigger the same generic error. In particular, union/anyOf-style JSON schema constructs tend to fail; union schema types are not supported in Information Extraction today.

Answers to your questions

• Are there known issues with saving Information Extraction Agents?
  Yes. We’ve seen intermittent failures tied to model serving/rate limits and incomplete error reporting; these surface as “Update Agent failed” even when the root cause is transient. There are also schema-related validation gaps (for example, unions) that currently fail with the same generic UI message.

• Does the dataset have to be a materialized table instead of a view?
  The docs specify using a Unity Catalog volume (folder) or a Unity Catalog table. Views aren’t mentioned/supported for the IE build flow, so use a table (or a volume), and if your inputs start as PDFs, convert them into a UC table first and point the agent at that table/column.

• Is there a way to see detailed logs for this failure?
  For the “Use PDFs” conversion flow, the UI links to a Databricks job—open the job to see failure messages. For the agent Save/Update itself, the UI currently doesn’t expose granular diagnostics; Support/Eng can inspect internal traces if you provide the agent URL and workspace ID (this is routinely requested on internal triage threads).

• Are there undocumented validation rules that may block saving (length limits, formatting, etc.)?
  The main ones to keep in mind:

  • Supported inputs: UC volume or UC table; labeled response must be a JSON string with the same keys in every row (for labeled datasets).

  • Supported schema types: string/int/float/boolean/enum/object/array. Union types (for example, anyOf with null or mixed types) are not supported and can cause “Update Agent failed.” Prefer explicit nullable handling at the field level (for example, allow empty string) rather than unions.

  • Context/size: IE agents have a 128k token max context length; extremely large examples/instructions can hit limits indirectly during save/update when samples are generated.

  • Workspace prerequisites that, if not satisfied, can block progress: Unity Catalog enabled, access to foundation models in UC (system.ai), serverless budget policy, compute/serverless requirements, and supported AWS regions (us-east-1 or us-west-2 for AWS).

Quick troubleshooting checklist

• Verify workspace prerequisites
  Ensure Unity Catalog is enabled, Agent Bricks (Preview/Beta) is enabled, a serverless budget policy exists, FM access is available via system.ai, and the workspace is in a supported region (AWS us-east-1/us-west-2).

• Use a UC table (not a view) or a UC volume
  If your source is PDFs, confirm you used “Use PDFs” to create a UC table, then point IE to the STRING column in that table.

• Simplify the schema first
  Avoid union constructs (for example, anyOf with ["string","null"] or mixed item types). Start with simple scalar types plus arrays/objects, and ensure labeled JSON (if used) has identical keys per row. Add fields incrementally to isolate the problematic field/type if the save fails.

• Reduce prompt/guidelines size temporarily
  Try saving with minimal global instructions/guidelines to rule out token-size issues, then re-introduce text gradually. IE agents have 128k context limits; extremely long guidelines can indirectly cause failures during sample generation.

• Retry after a short wait
  If this is rate limiting or a transient model-serving issue, a retry after a few minutes can succeed (this pattern has been observed; the UI unfortunately collapses various backend failures into the same message).

• Collect details for escalation if it persists
  Share the agent URL and workspace ID with Support so Eng can pull backend traces. This is the standard ask on triage threads and ES tickets to debug “Update Agent failed.”

Hope these tips are useful.

Cheers, Louis.

[Nadav]

Hey Louis,

Thanks a lot for your help! I implemented your suggestion and after replacing the base view I had created with a regular table, the issue was fully resolved.

Really appreciate you taking the time to look into it and share the insight!

Best,
Nadav