Agent block

An Agent is the LLM that does the reasoning. You give it a role and instructions, and attach the tools it may use. Every Flow has at least one Agent block. The block’s inspector configures:

1. General Name, description, system prompt
2. Model Which model to run and which inputs to accept
3. Data sources What it can read
4. Memory What it remembers
5. Retry What to do on failure
6. Rules How to clean up and post-process the model’s response

​

1. General

​

Basics and the prompt

Name The block name shown on the canvas. Description One line on what this Agent does. System prompt The instructions you give the Agent. Press the ✦ button at the top right to open an input box: describe the behavior you want in plain language, press [Generate], and it drafts the prompt for you. Long prompts expand with [Show more].

​

Cost

Turn on Max cost to cap what this Agent can spend in a single run.

​

Pause for approval

Pauses the run mid-way and continues only after a person confirms. All off by default.

• Before each turn Each turn requires approval before it starts.
• On routing decision Only available when the dispatch mode is Routed. When on, the run pauses after the router picks the next block, and an endpoint is provided for approving or rejecting the choice. Send the decision via the nora approvals CLI or the approval webhook and the run resumes.
• When cumulative cost exceeds Once spend passes the line you set, approval is required to continue.

​

Performance

• Bypass context cache When the Flow’s read-once context (Flow settings → Performance) is on, exempts just this Agent.

​

Dispatch mode

Appears only when the Agent has two or more wires going out to other Agent blocks. It decides how the next block is picked (Parallel · Conditional · Routed). See Dispatch mode in Wiring for how each mode behaves.

​

2. Model

​

Model selection

Mode and Model Choose where the model is called from, then select the appropriate model.

• API (LLM Provider) Calls an external provider’s API. The default. The list shows models from the LLM providers you have keyed into the workspace.
• Internal (On-prem GPU) Runs on GPUs you host in-house. Endpoints registered in Internal resources appear in the list. Use this when regulation keeps data from leaving your infrastructure.

Model-specific parameters The parameters the selected model supports can be configured here.

• Temperature Controls how random and creative answers are. Lower temperatures give consistent answers; higher ones give more creative, less predictable answers.
• Reasoning Appears for reasoning models. Choose how much effort the model spends reasoning before answering: Default, minimal, low, medium, or high.
• Language Appears for STT models. Sets the language of the audio to transcribe.

​

Inputs

The input kinds this Agent accepts. Off by default.

• File
• Image
• Voice
• OCR

​

Native tools

Tools the model runs server-side without a separate block. Select and enable the tools you need, and the Agent can use them while it runs. Web search, code interpreter, and file search are available depending on the selected model, and only supported tools appear in the inspector.

​

3. Data sources

Connect Knowledge, memory, and causal graphs that the Agent can read. The connection is per-Agent, not per-Flow, so different Agents in the same Flow can connect different sources. See Attach knowledge to Agent for detailed knowledge connection steps.

​

4. Memory

Connect a memory space so the Agent remembers what it learned between runs. The connection is per-Agent, not per-Flow: other Agents in the same Flow can use different memory. See Memory connection → Attach to an Agent for detailed connection steps.

​

5. Retry

Design what to do on failure. With no rules, the run fails immediately as before.

​

Limits

• Timeout (seconds) A single attempt over this long counts as failed.
• Circuit breaker (consecutive failures) Stop trying after this many failures in a row.
• Budget · total attempts The most attempts allowed in one run.
• Budget · total cost ($) The most one run may spend.

​

When exhausted

Pick what happens once the ceilings are spent.

• Fail Ends the run as failed. The default.
• Route to node Hands off to another block you selected in advance.
• Return partial Returns whatever was produced so far.
• Flag for review (keep going) Marks the run for review and continues.
• Notify a human Escalates for a person to look at.

​

Retry rules

Use the [+ Rule] button to add rules when different failures deserve different responses. Rules higher in the list apply first.

• Trigger Which failure this rule matches.
• Recovery point Where to retry from.
• Max attempts · total The most retries this rule gets.
• Backoff How the wait between retries grows.
• Delay (ms) How long to wait before retrying.
• Keep context When on, retries carry the earlier attempt’s context along.
• Attach failure cause When on, passes why it failed into the next attempt.
• Duplicate execution handling How already-executed tools are treated on retry. Prevents a hard-to-undo tool, like sending an email, from running twice.
  • Skip side-effect tools · recommended Tools not marked retry-safe are not run again on retry. An email or payment that already went out does not happen twice.
  • Require dedupe key Retries a tool call only when it carries a dedupe key. The tool is assumed to handle the same key idempotently, so it is run again.
  • Re-run everything (careful) Calls every tool again. Only for Flows you have explicitly confirmed safe.

​

6. Output rules

After the model responds, output rules clean up the result in predefined ways. They fix formatting rather than make a new judgment. Rules run from top to bottom, and each can be turned on or off. If there are no rules, the model’s output is used as-is. Use them when a model response contains a control token like <｜｜DSML｜｜>, wraps JSON in a code block, or returns a decision as text such as “approves it.” Output rules let you correct the result without rewriting the prompt.

​

Adding rules

Press the [+ Rule] button in the inspector to open a menu split in two.

• Quick templates Five templates provided by Nora.
  • Strip control token Removes <｜｜DSML｜｜> and similar model junk
  • Repair to JSON Strips ```json fences and prose, extracts valid JSON
  • Snap decision to enum Snaps to approve / reject or your allowed values
  • Collapse whitespace Regex \s{2,} → single space
  • Reassemble citation Reformats citations into a fixed field order
• Blank Pick just the kind and fill in the details yourself.

​

Rule kinds

Each kind expands the fields it needs (pattern, replacement, regex toggle, allowed values, and so on) inside the rule’s card.

​

Order and enabling

• Use the ↑ / ↓ arrows on each rule card to reorder. Higher rules run first, so lower rules see the already-cleaned output.
• The Enabled checkbox turns a rule off temporarily. Disabled rules are dimmed and skipped.
• Remove deletes the rule entirely.