Observability & Traces
Observability is what transforms an AI agent from a black box into a platform you can operate with confidence.
What you can inspect
Run Executions
Every message processed by any bot — time, channel, user, status, cost.
Tool Calls
Exactly which tools were called, with what arguments, and what was returned.
Latency Breakdown
Per-step latency: memory load time, model inference time, tool execution time.
Memory Reads & Writes
What was read from memory, what was written, and what was missing.
Error Details
Full error messages, stack traces for integration failures, model errors.
End-User Sessions
Filter by user ID to see all their runs — useful for per-user debugging.
Anatomy of a trace
Every run produces a structured trace you can open in the dashboard:1
Input
The raw user message, channel it came from, and the session ID.
2
Context loaded
What memory was injected — conversation history, persistent keys, user profile.
3
Model selected
Which model was routed to and why — based on agent config or dynamic routing rules.
4
Tool decision
If a tool was selected: which tool, the arguments sent, the raw response returned.
5
Response generated
The final output, token count, and cost.
6
Memory updated
Any memory keys that were written during this turn.
7
Run metadata
Total latency, status (success/error/timeout), channel, timestamp.
Debugging workflow
1
Find the failing run
Open Runs in your bot dashboard. Filter by status:
error, timeout, or tool_failure.2
Open the trace
Click any run to see the full trace tree. Expand each step to see input/output at that stage.
3
Identify the issue
Common issues:
- Wrong tool selected → Update instructions to be more specific
- Tool returned null → Check integration credentials
- Model hallucinated → Add explicit rules to instructions
- Context window overflow → Switch to
summarizedmemory
4
Use HiveMind Debug
Click Debug with HiveMind on any run. The AI will read the trace and suggest what to fix.
5
Fix and redeploy
Update your agent config or HiveLang file. Changes deploy instantly.
Bothive Observability vs LangSmith
LangSmith is an observability and evaluation tool built for LangChain apps. Bothive’s observability is built into the platform — no extra setup needed.
Observability in Swarms
When running multi-agent swarms, each agent’s trace is nested under the swarm run:Alerting
Set up run-level alerts in your bot settings:
Alerts can be delivered to Slack, email, or a webhook endpoint.