Platform/AI Observability
AI Observability

Can I see everything my AI is doing?

Full-spectrum visibility into every agent action, decision, and data flow — in real time.

Varman gives you a continuous, real-time window into every action your AI agents take — from the first token generated to the final tool invocation.

See Your Agents LiveBook a Demo

Live Trace Viewer

Watch your agents in real time. Every call, every decision, every millisecond.

Live Trace Viewer
AgentActionDurationStatus
ResearchAgentweb_search
312ms
success
SummaryAgentllm_invoke
847ms
success
EmailDraftAgenttool_call:send_email
94ms
blocked
DataAgentdb_query
156ms
success
ReviewAgentllm_invoke
623ms
warning
01

Execution Traces

Every agent action generates a structured trace — capturing the full call graph, tool invocations, model calls, latency at each step, and the complete input/output payload. Traces are stored immutably and indexed for instant search.

  • Full call-graph visualization
  • Sub-step timing breakdowns
  • Cross-agent trace stitching
  • Instant search across 90 days
TRACE: req_01HX9K2M → 847ms total
ResearchAgent.run()847ms
↳ web_search('Q4 market')312ms
↳ llm_invoke(claude-3-5-sonnet)489ms
↳ tool_call: summarize()34ms
↳ return_result()46ms
timestamp:2025-06-30T14:32:11.847Z
agent_id:"ResearchAgent-prod-47"
model:"claude-3-5-sonnet-20241022"
input_tokens:2847
output_tokens:412
latency_ms:489
cost_usd:0.00341
status:"success"
02

Structured Logs

Every model invocation emits structured, queryable logs capturing token counts, cost, latency, model version, and the full prompt/response pair. Filter, search, and export logs for debugging, cost attribution, and compliance.

  • Per-request token accounting
  • Cost attribution by team/agent
  • Structured JSON export
  • Real-time log streaming
03

Reasoning Inspection

For models that expose chain-of-thought, Varman captures and stores the full reasoning trace — every thought step, intermediate conclusion, and decision branch. Understand not just what your agent did, but why it did it.

  • Full chain-of-thought capture
  • Decision branch visualization
  • Reasoning diff across retries
  • Audit-ready reasoning export
REASONING CHAIN — ReviewAgent
1

I need to check if the output contains personally identifiable information before returning it to the user.

2

Scanning for: email patterns, phone numbers, SSN formats, credit card patterns...

3

Found 3 email addresses in the response. These must be redacted per policy DATA-PII-001.

4

Applying redaction. Original preserved in audit log. Returning sanitized output.

REPLAY: session_7f3a9b2c — 6 steps
Step 4 / 6 — llm_invoke
model: claude-3-5-sonnet · tokens: 2,847 · 489ms
Output: “Based on my analysis of Q4 trends...”
04

Session Replay

Rewind and replay any agent session step-by-step — scrub to any point in time, inspect the exact state at each step, and reproduce issues deterministically. Like a DVR for your AI agents.

  • Step-by-step scrubbing
  • State snapshot at each step
  • Shareable replay links
  • Deterministic reproduction
05

Audit Trail

Every action taken by every agent is written to an immutable, cryptographically-signed audit log. Query, export, and present audit data to compliance teams, auditors, and regulators with confidence.

  • Cryptographically signed entries
  • 90-day default retention
  • SOC 2 & ISO 27001 ready
  • One-click audit export
AUDIT LOG✓ SIGNED
14:32:11#a7f3b2c1llm_invoke[ResearchAgent]ALLOW
14:32:15#b8e4d3f2web_search[ResearchAgent]ALLOW
14:32:28#c9f5e4a3send_email[EmailAgent]BLOCK
14:32:31#d0a6f5b4db_query[DataAgent]ALLOW

Observability Outcomes

0%
Agent action visibility
Every call traced
<0ms
Trace latency
Real-time by default
0 days
Audit retention
Default, configurable to 7yr
0
Blind spots
Across all agents

See Your Agents Live

Connect Varman to your AI infrastructure in minutes. Full observability, zero code changes required.

Book a DemoBack to Platform