AI agent investigations
Start with the evidence you actually have.
Understand which actions were checked, which policy applied, and where the record stops. Investigation starts with a clear distinction between a decision and an observed outcome.
Try before connecting
The public sample uses synthetic records: an allowed read with a host-reported outcome, a blocked settings change, and an allowed action without an execution receipt. It requires no account and reads no customer data.
Keep your existing controls
Investigation is read-only for existing Pramagent records; you do not need to install a new hook to read them. A workspace owner can invite an Investigator. Forensics MCP reads collected evidence; it does not import logs. The private CLI has bounded offline import adapters, but the hosted console has no general log-upload onboarding. Imported events remain source claims; signing at import cannot prove what happened before collection. Ask the operator about CLI availability; do not assume it is in a public package.
Choose future capture explicitly
Workspace owners and admins can revisit Evidence settings after setup. Metadata-only preserves decision digests, not raw tool arguments. Customer-retained mode leaves matching inputs under your control. Encrypted snapshots require operator-configured storage, capture policy and workspace enrollment. The console shows the effective mode. Changing it neither recovers missing old content nor deletes previous snapshots; approvals, imports, exports and backups remain separate retention domains.
Reconstruct a session
Filter recorded decisions by session and time. Inspect tool calls, policy digests, approval references, and fingerprints where available. Sequence edges show order; explicit parent-span references show recorded relationships. Older records may lack these fields, and the result identifies its selection limits.
Replay the policy decision
New supported traces bind a content-addressed policy snapshot and engine version into the audit record. Replay re-evaluates supported rules without running tools. Exact matching input and required state must be available. Missing evidence produces a partial or unavailable result, never a claim of successful reproduction. Consistency checking remains a separate operation.
Test a candidate rule
Retro-hunt evaluates a candidate rule over selected past records and reports changed decisions and unevaluable records. Missing original content cannot be recovered from its hash. A hunt does not execute historical actions or change your live policy.
Preserve a review copy
Download a signed ZIP containing selected records, policy snapshots, chain material, a manifest, and a standalone verifier. Pin the exporter public key through a trusted channel to check the signature offline. Available evidence-v2 material can be verified with its own trusted keys. Export content depends on the original trace; review it before sharing.
What verification establishes
Verification detects changes to the signed export and checks the included evidence. It does not prove that every event was collected, independently timestamp an event, identify an attacker, or establish that a permitted tool ran. Independent checkpoints and execution receipts only strengthen these claims when actually present and verified.
Keep conclusions separate
Repeated requests can reflect retries or integration behavior. They do not by themselves prove malicious intent. Pramagent currently provides deterministic evidence views, not an AI-generated attribution or legally certified forensic report.