Agent mode: will per-turn context reporting be added, or is there another way to read it?

ncastagnet_mc — Fri, 28 Aug 2026 18:22:01 GMT

Genie spaces are steered by a semantic layer: text instructions (e.g. "when the user says 'sales', always filter to region = 'EMEA'") and example question→SQL pairs.

We read turn payloads programmatically — via the message endpoint behind the Genie UI (/api/2.0/data-rooms/{space}/conversations/{cid}/messages/{mid}), with the public conversation API as fallback (it's thinner and omits exactly the context below).

Comparing the two modes:

Standard mode tells you what each turn used. Every message payload records the instructions applied (progress_report.context_info.primitive_instructions), the example SQL pairs it drew on (examples.example_questions, with confidence scores), and the tables retrieval selected (relevant_tables). When an answer looks wrong, you can check whether the instruction fired before touching any SQL.
Agent mode reports none of this. The payload carries only the research trail — sub-questions, research thoughts, queries, the final report. Nothing records which instructions or example SQL shaped the turn. And this is the mode where it matters most: one deep-research turn can run 5+ SQL queries, so a missed instruction silently propagates across all of them.

Specifically, what we need: a new output item type in the agent mode API's response stream (POST /api/2.0/genie/agents/{agent_id}/responses) carrying the per-turn context standard mode already reports: primitive_instructions (instructions applied), example_questions (example SQL used, with confidence), relevant_tables (tables selected).

The space API doesn't close the gap. It exposes the configured instruction/example list — what could have steered a turn, not what did — and it's current-state only: no version history, so after an edit you can't reconstruct the config a past turn ran against. (If config versioning is planned, that would help here too.)

Questions:

Will agent mode report per-turn context (instructions / example SQL used) the way standard mode does? If not, is it on the roadmap?
Is there another way to read it today — an API field, an event type in the stream, a debug option, or the UI?

Thanks!

Re: Agent mode: will per-turn context reporting be added, or is there another way to read it?

DoTA — Sun, 30 Aug 2026 00:14:31 GMT

Checked the current Agent mode Responses API docs on this (docs.databricks.com/aws/en/genie-agents/api) and your read looks right, not a gap in your search.

The response stream only documents four output item types: reasoning, function_call, function_call_output, and message. None of them carry text_instructions, example_question_sqls, or relevant_tables the way progress_report.context_info does in standard mode. So as things stand there's no documented field or event type that reports per-turn context in Agent mode.

One partial workaround: function_call items expose the literal execute_sql arguments, i.e. the SQL that actually ran. You can diff that against the space's configured example_question_sqls and table list (from serialized_space) to infer which example/table likely influenced a query. That's inference after the fact though, not a record of which instruction fired, so it won't catch your "missed instruction silently propagates across 5+ queries" failure mode.

Given Agent mode APIs are still Beta, this seems like exactly the kind of gap the Genie team would want flagged directly, e.g. via the in-product feedback widget or a support ticket linking this thread, since whether it's roadmapped isn't something the docs will answer. Tagging in case anyone from the Genie team is watching this board and can confirm either way.

topic Agent mode: will per-turn context reporting be added, or is there another way to read it? in Generative AI

Agent mode: will per-turn context reporting be added, or is there another way to read it?

Re: Agent mode: will per-turn context reporting be added, or is there another way to read it?