Skip to main content

Debugging a session

When an agent does something inexplicable, the debug view shows what it was actually working with — as opposed to what you think you configured.

Who can do what​

TaskUserAgent ownerAdmin
Open the debug view——✅

Agent → Debug, for a specific session.

What it shows​

SectionWhy it matters
System promptEverything the agent was told before your message — organization bootstrap, agent bootstrap, identity, memories, tool instructions. This is where "why is it doing that?" is usually answered.
ToolsThe exact tool list the agent had, with a count. Compare against what you expected.
Session status & task stateWhere the task got to, and the scratchpad it kept through a long job.
MessagesThe raw sequence, including tool calls and their results.

How to use it​

"It ignored my instruction." Read the system prompt. Instructions arrive from several places — organization default bootstrap, the agent's own bootstrap, identity, memories — and they can contradict each other. The one that lost is usually visible.

"It didn't use the tool I gave it." Check the Tools section first. If the tool isn't there, it's a configuration problem — an unattached bundle, or a missing connection. If it is there and went unused, the tool's description didn't match the request.

"It didn't use the skill." Same shape: the skill's description determines whether it triggers. See Skills: authoring.

"It forgot something from earlier." Look at task state and the message sequence. Long tasks compress history; anything that must survive belongs in memory or task state, not in conversation.

"It stopped mid-task." Check the session status. If a budget stopped it, the enforcement audit log names the limit.

Before you reach for it​

Cheaper checks first, in order:

  1. Is the agent Active? (Agent → Settings → General)
  2. Does it have the tool, and is the underlying service connected?
  3. Did a budget stop it?
  4. Does the model it's pinned to still exist and is it still enabled?

Most reports resolve at one of those four.

A note on what you'll see​

The system prompt includes memories and identity content, which can contain things members told an agent in confidence. Debug access is admin-only for that reason — treat what you read there the way you'd treat any other member data.