# When the Agent stops or asks

> Why a job stops, what each message means and what to do next. It stops rather than guess, and never retries a CAD step on its own.

Status: Available · Level: Deeper · About 6 minutes · Canonical: https://paracrux.com/docs/when-the-agent-stops

## You need

- You have run at least one Agent job

## You'll come away with

- Tell a refused request from an uncertain one
- Know what to check before you continue
- Know the limits that end a job

## Step by step

1. **It asks when something is missing.** If a size, unit or requirement is missing or ambiguous, the Agent is told to ask rather than assume. Answer and send again.
2. **Rejected: nothing happened.** If a request fails its checks before any file or SOLIDWORKS call, it is recorded as failed and you can simply try again with corrected inputs.
3. **Uncertain: check before you continue.** If a CAD step may have had side effects, the job stops and says so. Open the action's record and the output folder, check what exists, and decide yourself. The Agent will not try another CAD action in that message.
4. **Limits end a job.** A job ends at eight steps or its £2 ceiling, and each message allows at most 16 tool calls. Send another message to continue.
5. **If the app closes mid-job.** On restart, anything that was running is marked uncertain. Check it as above; nothing is replayed.

## Walkthrough

Each message is quoted from the app. The situations that trigger them are examples. The routing frames are recorded benchmark answers from a layer that runs in shadow mode.

- **You:** Open an existing part, make the lower hole 5 mm bigger in diameter, then produce a standard ISO drawing with a section view.
- **Route (shadow mode):** Leaning = unsupported; Confidence = 0.64, 0.54; Floor = 0.80; Result = Abstained: you decide
- *A recorded benchmark case. Below the floor, the routing layer abstains instead of choosing.*
- Provenance: recorded

- **Agent:** CAD action outcome is uncertain; inspect its recorded evidence before continuing.
- *Shown when a CAD step may have had side effects.*
- Provenance: real-code

- **Agent:** CAD operation already attempted or a previous action is uncertain. Inspect the action record; no automatic replay.
- *Shown if another CAD action is attempted in the same message.*
- Provenance: real-code

- **Agent:** Tool limit reached for this message; send another message to continue
- Provenance: real-code

- **Agent:** App stopped during dispatch; provider usage may have occurred. No automatic retry.
- **Gate:** Check the action record and the output folder, then decide what to do next.
- *Shown after a restart for anything that was running.*
- Provenance: real-code

## What it won't do

- An uncertain step may still have used credit; its hold is kept until usage is known.
- The routing layer shown here runs in shadow mode. It records its choice but does not steer the Agent yet.

## Covers

- Job limits (Available): One tool call per step, at most eight steps and £2 per job; credit is reserved per step and unused credit returned. Where: Every managed AI job
- Typed request routing (Lab): A separate model picks a fixed route for each request, or abstains below 80% confidence. Shadow mode only. Where: How it decides page

## Next

- [Get help from support](https://paracrux.com/docs/get-help.md): Open a request from your account or the app, choose what to attach, and follow replies. It works with or without an active subscription.