Skip to content
TapPass

Watch mode: policy from behavioral evidence

Guessing an agent’s policy up front is hard, and over-tight policy that blocks real work gets disabled fast. Watch mode inverts the problem: run the agent with nothing blocking, let TapPass record everything it does, then derive a policy from the evidence.

1. Watch by default

Section titled “1. Watch by default”

New agents start in shadow (watch) mode — every LLM call, tool call, and write is governed and recorded, but no verdict blocks anything. Your agent behaves exactly as it did before TapPass; you just start accumulating an evidence trail.

The same is true when you apply a policy template or create a policy: it lands in shadow first. Watch, then enforce.

2. See what the agent actually did

Section titled “2. See what the agent actually did”

The activity cube is the aggregated, pivotable view of that recorded behavior:

Terminal window
tappass activity cube -g tool                       # calls grouped by tool
tappass activity cube -g tool,param_key -f tool=gmail.send
tappass activity cube -g provider --since 30d --sort cost
tool                 calls  allowed  blocked  tokens   cost    last_seen
gmail.send             142      142        0       —    —       2m ago
github.get_file         88       88        0       —    —       5m ago
stripe.create_charge     3        3        0       —    —       1d ago

This is the ground truth for policy: gmail.send and github.get_file are core to the job; that lone stripe.create_charge is worth a hard look before it becomes normal. Stream it live with tappass session watch, and use tappass audit replay <event_id> to re-decide any past event against a candidate policy — a what-if with no side effects.

3. Derive the policy from behavior

Section titled “3. Derive the policy from behavior”

Instead of describing the agent’s job in words, let its observed activity write the envelope. In the dashboard, open the agent’s activity view and choose Propose policy: TapPass reads the lookback window you pick (from an hour up to months), plans the smallest capability envelope that covers what the agent actually did, and shows you the rules. On confirmation they land in shadow — ready to promote to enforce.

Prefer the terminal? tappass policy create --from-prompt is the description-first equivalent, and tappass audit replay lets you test any candidate against history before promoting.

Why watch mode

Section titled “Why watch mode”
  • Right-sized, not aspirational. The policy matches real usage, so it blocks anomalies without breaking daily work.
  • Evidence you can show. Every rule traces back to observed activity in the audit trail.
  • Safe promotion. You’ve already seen exactly what would have been blocked before you flip to enforce.

Prefer to start from intent instead of history? Govern an agent from a sentence is the description-first counterpart — both land the same kind of policy.

Next steps

Section titled “Next steps”