On this page
01
What you'll set up
An agent run is dozens of requests. When one goes wrong you need to see what it did, in order, with the tools it called and what came back. In about eleven minutes you will choose how much to keep, turn it on, and read a session.
- A retention mode that fits your privacy needs
- Session traces recorded for the policy that governs your agents
- A step-by-step view of a session
- Verification that each record is intact
You need an owner or admin role and a policy bound to your agent's traffic.
02
Choose how much to keep
Retention is a policy setting. Start with the least you need.
| Retention mode | What is kept | Use it when |
|---|---|---|
| Zero retention (tokens only) | Counts and cost. No content | You want no content stored at all. This is the default |
| Attribution only (headers + metrics) | Labels and metrics, plus tool names | You want to see which tools ran, not their content |
| Tool calls (arguments + results) | Each tool call, its arguments, and its result | You need to see what an agent did with its tools |
| Full audit (complete payloads) | The text of every step, as well as tool activity | You need a complete record for a regulated workflow |
03
Turn it on
Retention mode is on the Advanced step of the policy form.
- 1
Open the policy
Go to AI → Policies and open the policy that governs your agents.
- 2
Open Advanced
Find Retention mode, next to the throughput limits.
- 3
Choose a mode
Start with Tool calls (arguments + results) for most agents.
- 4
Save
New sessions are recorded from this point.
Throughput limits
- 3Retention mode
- Tool calls (arguments + results)
04
What gets a record
Records are made for steps that have something worth keeping.
| Mode | Steps with tool activity | Steps without |
|---|---|---|
| Zero retention | No record | No record |
| Attribution only | A record with tool names | No record |
| Tool calls | A record with arguments and results | No record |
| Full audit | A record | A record with the text |
Every request still appears in the session view from the ledger, in order, with its model, cost, and status. Records add the content on top.
05
Open the Sessions tab
Sessions lists agent runs, newest first.
| Column | Meaning |
|---|---|
| Session | The session id your agent declared or Cloptima inferred |
| Apps | The apps that took part |
| Requests | How many requests the session made |
| Recorded | How many have a trace record |
| With tool calls | How many steps called tools |
| Spend | What the session cost |
| Last activity | When it last ran |
Filter by app, environment, or team to find the run you want.
06
Read a session step by step
Open a session to see Session Execution Steps.
| Column | Meaning |
|---|---|
| # | The order of the step |
| Time | When it ran |
| Model | Which model answered |
| Tools called | The tools the model asked for |
| Tokens and Cost | What the step used and cost |
| Trace | Whether a record exists and whether it is intact |
- 1
Expand a step
A step with a record shows each tool call, its arguments, and the result from the next step.
- 2
Follow a tool call
Each call shows as Requested, Result submitted, or Unresolved.
- 3
Copy ids
Request and call ids copy with a click, for tickets and searches.
Session Execution Steps
| # | Time | Model | Tools called | Tokens | Cost | Trace |
|---|---|---|---|---|---|---|
| 1 | 10:02:14 | claude-sonnet-5-5 | search_orders | 2,140 | $0.012 | Verified |
| 2 | 10:02:17 | claude-sonnet-5-5 | none | 2,986 | $0.016 | Verified |
1Step 1 · search_orders
| Tool call | Result submitted (step 2 returns it) |
| Arguments | { "order_id": "8841" } |
| Result | { "status": "shipped" } |
07
Verify a record
A trace is only useful if you can trust it. Each record carries a fingerprint, and the session view checks it every time you open it.
| Badge | Meaning |
|---|---|
| Verified | The record is intact and matches the ledger |
| Pending | The record is on its way to storage and will appear shortly |
| Missing | The ledger shows a record that could not be found |
| Hash mismatch | The stored record does not match its fingerprint |
08
Large values are shortened, not dropped
Some tool arguments are enormous, such as a whole file being written.
Very large values are shortened to a readable head, and the record keeps the original size and a fingerprint of the full value. A reviewer can check a separately supplied original against it.
09
Privacy and retention
Keep what you need, for as long as you need it.
- Zero retention is the default, so nothing is stored until you choose
- Guardrails run before storage, so a redacted secret is not recorded
- Records are kept for the retention period of your plan, alongside the ledger
- Use Tool calls rather than Full audit unless you need the text of every step
10
If something goes wrong
Most questions are about which steps were recorded.
| What you see | Likely cause | Fix |
|---|---|---|
| No records at all | Retention mode is Zero retention | Choose a mode that keeps content |
| Steps without a record | They had no tool activity, and the mode keeps tool steps only | Use Full audit if you need every step |
| A step shows Pending | The record is still being written | Refresh in a few minutes |
| A cached answer has no record | It was served from cache, so no provider call happened | Expected |