InficySee what your AI agents did. Keep evidence you can verify.
Inficy turns captured AI-agent runs into readable, tamper-evident execution records. Inspect recorded steps, tool calls, failures, handoffs and human interventions, export the original evidence bundle, and optionally certify eligible runtime-observed executions.
Runtime-observed evidence · Agent-reported ChatGPT sessions clearly separated · Optional certification
Your model-provider keys stay in your runtime. Inficy captures execution evidence through supported connectors without proxying your OpenAI, Anthropic or xAI traffic.
example-research-agent
Completed after recovery
- 2m 51s
- 18
- 1
- Read 3 supplier websitesweb · 3 domains · 41s
- Called search.querytool · 4 calls · 6.2s
- Handed off to pricing-agentdelegation · returned in 22s
- Extraction failedparser error · retried once
- Operator answered a questionhuman · 1m 12s wait
- Renewal quote draftedoutcome: completed · exported
Your agents work out of sight. Inficy keeps what the capture integration observed.
Where it went
Tools, services, models and websites recorded by the capture integration, shown in execution order.
How long it took
Wall-clock duration for the run, and a separate breakdown of where the activity time went.
What happened
Completions, failures, interruptions, retries, recoveries and every point a human stepped in.
See where agent activity went
Inficy projects readable timelines and analytics from the stored execution evidence. The original NexArt Project Bundle remains authoritative. Wall-clock duration stays separate from recorded activity time, and activity the capture integration did not observe is never invented.
Runtime-observed dashboard metrics are kept separate from agent-reported ChatGPT session analytics.
Execution activity
- Executions
- 1,284
- Actions recorded
- 22,910
- Agents reporting
- 6
Time by activity
- Tool calls
- 38%
- Model calls
- 27%
- Waiting on humans
- 19%
- Web and retrieval
- 16%
Tools and services
- accounting.invoices
- 412 calls
- crm.contacts
- 268 calls
- mailer.send
- 97 calls
Websites and domains
- docs.supplier.com
- 184 requests
- api.exchange.io
- 96 requests
- status.provider.net
- 31 requests
Financial activity
- Requested
- 48
- Approved
- 41
- Succeeded
- 37
- Failed / refunded
- 3 / 1
Human interaction
- Approvals granted
- 63
- Approvals declined
- 7
- Median wait
- 1m 34s
Handoffs
- Agent to agent
- 112
- Agent to human
- 70
- Returned incomplete
- 9
Failures and recovery
- Failed actions
- 184
- Recovered by retry
- 141
- Interrupted runs
- 12
Outcomes
- Completed
- 1,097
- Completed after recovery
- 141
- Failed or interrupted
- 46
Open any execution and read it like a story
- Every recorded step with its timestamp, type, duration and outcome, shown in sequence.
- Errors and interruptions stay in the timeline instead of being filtered out.
- Approvals, declines and manual retries marked where they happened.
- Payment states exactly as observed: requested, attempted, approved, succeeded, failed or refunded.
Tamper-evident evidence, not a mutable activity log
Tamper-evident
Each completed runtime-observed execution is preserved as a sealed NexArt Project Bundle, so later alterations are detectable.
Workspace-isolated
Records belong to your workspace only, enforced at the database level.
Exportable
Take an execution, or a filtered set of them, out as structured evidence.
With confidential capture, selected sensitive values are replaced by NexArt commitments before the artifact leaves your runtime. Private opening material stays with you.
Need stronger evidence for one execution?
When an eligible runtime-observed execution needs to be shared or independently checked, submit its original stored NexArt Project Bundle for certification. Certification is optional, applies to one selected execution and never changes or reconstructs the recorded artifact.
Certification attests the stored artifact at certification time. It does not prove that the agent was correct, safe, complete or compliant.
How certification worksStart free, upgrade when you scale
Recording is included on every workspace. Starter and Growth add higher allowances, longer retention and included certification credits.
See plans and what each one includesEvidence, audit trails and the limits of proof
What an execution record should contain, what an agent audit trail needs, and the difference between capture, verification and independent certification.
Read the blogSee the execution record, not just the result.
Connect a supported runtime or add the ChatGPT connector. After the connection is verified, recorded activity begins appearing in Inficy.