Traces. Capture, share, and learn from coding agent sessions.
Traces captures coding agent sessions across the team, and makes them readable, and searchable. Traces are linked to the git commits, files and other artifacts they produced, helping teams access their collective memory and learn from their work with AI.
We helped make Traces a more reliable platform for capturing agent work and turning it into knowledge teams and their agents can build on.
We improved the modelling of traces, supporting subagents, compaction, and captured most AI harness features into traces. We redesigned trace sharing, making it 5–15× faster while sending less data in our internal benchmarks. We also improved resilience of sharing as we scaled up to 100k+ traces.
We strengthened the foundations of session capture with stable identifiers for traces, messages, and parts, more reliable git and agent hooks, and more complete capture of agent sessions across adapters. We also helped build bulk uploads of local traces to make onboarding easier for new users.
We built tools to retrieve information from traces, and an agent around it that answers questions about traces. This was then further extended to an MCP that people could access from their own coding agents.
We helped build Surfaces, letting users explore their agent sessions from different perspectives, including file changes, security reviews, token usage, and skills. We designed Surfaces so teams could create and publish their own views.
Lastly, we built the infrastructure to report over 100k+ traces for an organization, creating dashboards for CISOs, CFOs etc, that provide insights on AI usage and coding agent patterns over many months of work with agents. This provided the foundation for Trace Intelligence.