Agent sessions

The Agent sessions report brings together coding-agent conversations, evaluations, and linked delivery work. Use it to understand how agents approach tasks and what those sessions contribute to your codebase. The report is powered by AI Code Insights.

Note: This report requires AI Code Insights, Transcripts & Agent Experience enabled, and permission to view individual metrics in reports. Transcript access is controlled separately.

When to use Agent sessions

This report helps you:

  • Explore agent work — Find sessions by contributor, tool, or work type.
  • Understand an interaction — Review the conversation, tool calls, and evaluation behind a session.
  • Follow work through delivery — Trace a session’s contributions to linked issues, PRs, and deployments.

All sessions report

The chart shows session volume over time. The list below it shows individual sessions, including their contributor, title, work types, Agent Experience Score, tool, message count, and date.

Agent sessions report showing session volume, filters, and the session list

Use the date, team or group, attribute, Tools, and Work types filters to narrow the report. Search for a contributor, email, tool, or session title, then select a session to inspect its details.

DX counts each top-level session once, using its start time. Sessions can appear without a linked PR or evaluation score. Captured subagent tasks appear within their parent session rather than as separate rows.

Session details

Session details connect the conversation to its evaluation and delivery context. You can open a session from this report or from session drilldowns in Agent effectiveness.

Agent session details showing the summary, work lineage, and Agent Experience Score

Overview

The Overview tab summarizes the task, contributor, agent, model, work types, and complexity, alongside timing, message counts, and token usage. Duration measures elapsed time from session start to the last recorded message, including idle time.

Use the sections below the summary to investigate the session:

  • Work lineage shows the issues, PRs, and deployments associated with a session. See how DX links work to sessions.
  • Agent Experience Score shows ratings for requirements, steering, and scope, with comments explaining the evaluation. See the scoring rubric.
  • Model fit (beta) compares task complexity with the model used and suggests alternatives where available. See Model fit.
  • Tools used in this session shows reported Skill and MCP usage, grouped by Skill name or connected service, with invocation counts.

Available information depends on the agent integration, evaluation results, and your transcript permissions.

Activity

The Activity tab shows the conversation alongside captured tool calls, helping you see how the agent approached the task and responded to guidance. Messages include timestamps, models, and token counts when available. Expand a tool-call group to inspect individual calls, their status, and duration.

Search messages or filter by Type to find a specific exchange. Use Copy direct link to share a message. CSV exports respect your search and type filters; when opened from a message link, the export starts at that message. See Transcript handling for collection and scrubbing details.

Subagent tasks

Captured delegated work appears as a Subagent task within the parent conversation. Select View to inspect the exchange between the parent agent and subagent.

Subagent relationships describe delegation; Work lineage describes delivery connections. Subagent capture depends on the agent integration.

Agent work

Agent work groups transcript-powered sessions by work type and category. A session can receive up to three work types, so work-type counts are not mutually exclusive and should not be added together as a session total.

Work types appear in the session summary and can be used to filter the Agent sessions report. Use them to check whether friction clusters around testing, refactoring, feature work, or another task type. For trends across sessions, see Agent effectiveness.

Agent work classifies sessions into these categories:

  • Exploration — Questions, ideation, architecture design and planning
  • Features — New features, feature enhancements
  • Fixes — Bug fixes, build or test failure fixes, data or migration fixes, vulnerability remediation and hardening
  • Maintenance — Refactors and rewrites, dependency or config updates, test updates, documentation, performance improvements
  • Code review — Code reviews
  • Other — Data generation, other

Model fit (beta)

Note: Model fit requires AI Code Insights with the daemon and Transcripts & Agent Experience enabled.

Model fit compares the assessed complexity of a session with the model used for the work. It helps you see when a lower-cost model may be enough or when a more capable model may improve results.

DX assesses task complexity from the transcript and uses the primary model recorded for the session, then compares the task’s complexity with the model’s recommended complexity. In a session’s Overview, the result is:

  • Well matched — The model’s recommended complexity matches the task’s complexity.
  • Potential savings — The model is recommended for more complex work, so a lower-cost model may be enough.
  • Potential quality improvement — The model is recommended for less complex work, so a more capable model may improve results.
  • Not assessed — Task complexity could not be determined or the model is not supported.

When available, DX suggests an alternative model from the same model family that matches the assessed complexity. These are opportunities to investigate, not measured cost savings or proof of output quality.

Model fit across sessions

The Agent effectiveness report groups Model fit results by work type and model. Select a model name to sort its opportunities, or select Work type to sort by name. Select a result to see the sessions behind that work type and model pairing.

A pairing is Well matched when neither a savings nor a quality opportunity applies to 25% or more of its assessed sessions. Otherwise, it shows the more common opportunity; if both have the same share, it shows Potential savings. A pairing with no assessed sessions is Not assessed. Model columns with no assessed sessions are hidden.

The assessed-session count includes sessions with both a task complexity rating and a supported model. A session can have up to three work types, so it may appear in more than one row. Individual session results can differ from the aggregate result for a pairing.

Model fit currently supports:

  • Claude Fable 5
  • Claude Fable 5.1
  • Claude Haiku 4.5
  • Claude Haiku 5.5
  • Claude Opus 4.5
  • Claude Opus 4.6
  • Claude Opus 4.7
  • Claude Opus 4.8
  • Claude Opus 5
  • Claude Opus 5.5
  • Claude Sonnet 4
  • Claude Sonnet 4.5
  • Claude Sonnet 4.6
  • Claude Sonnet 5
  • Claude Sonnet 5.5
  • Gemini 2.5 Flash
  • Gemini 3 Flash
  • Gemini 3 Pro Preview
  • Gemini 3.1 Pro
  • Gemini 3.5 Flash
  • Gemini 3.6 Flash
  • GPT-5 mini
  • GPT-5.1
  • GPT-5.2
  • GPT-5.3 Codex
  • GPT-5.4
  • GPT-5.4 mini
  • GPT-5.4 Nano
  • GPT-5.5
  • GPT-5.6 Luna
  • GPT-5.6 Sol
  • GPT-5.6 Terra
  • GPT-6 Astra
  • GPT-6 Luna
  • GPT-6 Sol
  • GPT-6.1 Sol
  • Grok 4.5
  • Grok 4.6
  • Grok 4.7
  • Kimi K2.7 Code
  • MAI-Code-1-Flash

Sessions that use other models are not assessed for Model fit. Contact your DX account team if you need another model supported.

Work lineage

Work lineage shows how a session connects to the work around it: related issues, a repository and branch, PRs, and deployments. Open an issue or PR card to inspect it in the source system, or expand a stack to reveal additional items.

DX builds the diagram from recorded code attribution and imported delivery data:

  1. Session to commits: DX finds committed files attributed to the session and identifies their commits.
  2. Commits to PRs: DX matches those commits to PRs by commit SHA.
  3. PRs to issues: DX adds issues linked to those PRs and their immediate parent issues.
  4. PRs to deployments: DX finds successful deployments linked to those PRs that match your account’s deployment environment filter.

The Repo & branch shown in the session comes from the most recently created linked PR with repository and branch information. DX derives this from delivery data rather than the session’s working directory.

The deployment card shows the latest matching successful deployment’s environment and time, plus a count of additional deployments. The environment filter defaults to names containing prod, ignoring case. PRs can appear before they are merged or deployed.

Understanding session contributions

One session can contribute to several PRs, and one PR can combine work from several sessions and people. A PR appearing in a session’s lineage does not mean that session produced the entire PR. DX finds issues through their associated PRs, so an issue in the diagram does not necessarily represent the task that initiated the session.

No linked work means DX has not found a PR through the session’s recorded commit attribution. Sessions that produce no committed code can still have a conversation, summary, and evaluation. For sessions that did produce code, check that commits and source control data have reached DX. Issues and deployments appear only when DX has associated them with those PRs.

Access and configuration

Admins enable Transcripts & Agent Experience under Admin → AI Code Insights. Two settings in Access & visibility control what you can see:

  • Individual metrics visibility → Reports controls report access and team scope.
  • AI Code Insights transcripts controls recorded titles, summaries, evaluation comments, and Activity. Without access, session details use a generic tool-and-date title and hide those fields.

A session opened from another report can remain accessible when the Agent sessions report is hidden. Transcript permissions still apply to shared links. See Individual contributor metrics for permission options.