Agents

​

What Are Agents?

• Adaptive: Different approaches for different situations
• Tool-driven: Use integrations and workflows as needed
• Conversational: Interactive back-and-forth discussions
• Context-aware: Consider your instructions and conversation history

​

How Agents Work

​

Tools (What Agents Can Use)

​

Instructions (How Agents Behave)

When the user provides a ticket number:
1. Retrieve ticket details and ask for confirmation
2. Once confirmed, run the "Get User Profile" workflow
3. Read the discount policy from Notion
4. Assess eligibility and explain your reasoning

​

Reasoning (How Agents Think)

1. Analyzes the request and available tools
2. Decides which tools to use and in what order
3. Executes tool calls and adapts based on results
4. Asks for confirmation when needed (based on instructions)
5. Explains its reasoning step-by-step

​

Creating Your First Agent

You're a support operations assistant helping evaluate discount eligibility.

When given a Zendesk ticket number:
1. Retrieve ticket details and display them
2. Ask for confirmation before proceeding
3. Run "Get User Profile" workflow
4. Read discount criteria from Notion page [URL]
5. Evaluate eligibility with clear reasoning

Always respond professionally. If uncertain, ask for clarification.

Limiting Tool Capabilities (Recommended)

For MCP integrations, you can restrict which specific tools the agent can access. This makes your agent more reliable and prevents unintended actions.How to limit tools:

1. Click on any MCP integration you’ve added (Gmail, Salesforce, Slack, etc.)
2. Toggle off specific tools you don’t want the agent to use

Example configurations:

• Gmail: ✅ Search emails, Read emails → ❌ Send email, Delete email
• Salesforce: ✅ Get account, Search records → ❌ Delete record, Update record
• Slack: ✅ Read messages, Search channels → ❌ Send message, Create channel

Why this matters:

• Prevents destructive actions (deleting, sending)
• Makes agent behavior more predictable
• Reduces risk of unintended operations
• Improves reliability by limiting decision space

Add this to your system prompt: “If a user asks you to perform an action that’s disabled (like sending an email), explicitly tell them that capability is not available for this agent.”

​

Embedding Agents in Workflows

​

Code Sandbox Tool

​

What is the Code Sandbox Tool?

​

Available Operations

​

Example Use Cases

"Upload my sales_data.csv file and create a summary report with:
- Total revenue by month
- Top 10 products by sales volume
- A bar chart showing monthly trends"

"Take this PDF report and extract all the tables into a CSV file"

"Calculate the compound annual growth rate (CAGR) for each product 
in my portfolio over the last 5 years and rank them"

​

Using Workflows as Agent Tools

​

Designing Workflows for Agents

Input Node: "company_name" (text)
Input Node: "date_range" (text)
[Your workflow logic here]
Output Node: "enriched_data" (object)
Output Node: "success" (boolean)

​

AI Advanced Settings

• Model Config
• Auto Summarization
• Fallback

Customize AI model behavior. Available parameters vary by provider (OpenAI, Anthropic, Google). Settings are stored per-provider, so switching models preserves your preferences for each.

OpenAI Parameters

Parameter	Range	Default	Description
Reasoning Effort	low, medium, high	medium	Controls computational effort before responding. Higher = more thorough reasoning but slower and more tokens. Only for o-series and GPT-5 models.
Temperature	0–2	1	Controls output randomness. 0 = deterministic, 2 = highly creative.
Max Output Tokens	1+	Auto	Upper bound for generated tokens (includes reasoning tokens). If unset, uses model default.
Max Tool Calls	1+	Auto	Limits total tool calls per response. Useful for cost control.
Top P	0–1	1	Nucleus sampling. Model considers only tokens in the top P probability mass. Only for GPT-4o, GPT-4.1, GPT-5.1, GPT-5.2, o3, o4.
Parallel Tool Calls	on/off	on	When on, model can execute multiple tools simultaneously. Disable if tools depend on each other’s results.

Adjust either Temperature or Top P, not both. Using both produces unpredictable results.

Anthropic (Claude) Parameters

Parameter	Range	Default	Description
Extended Thinking	enabled, disabled	enabled	Shows Claude’s reasoning process before the final answer. When enabled, temperature is forced to 1.0.
Budget Tokens	1,024+	10,000	Token budget for thinking. Higher = more thorough reasoning but slower. Only shown when thinking is enabled. Must be less than Max Tokens.
Temperature	0–1	1	Controls output randomness. Note: forced to 1.0 when Extended Thinking is enabled.
Max Tokens	1+	Auto	Maximum tokens to generate. Defaults: 64,000 (Claude 4.5), 8,192 (older models).
Top P	0–1	1	Nucleus sampling threshold.
Top K	1+	Auto	Limits sampling to top K most likely tokens. Lower values (10–40) produce more focused outputs by removing long-tail responses.
Disable Parallel Tool Use	on/off	off	When on, model outputs at most one tool call per response. Enable for strict sequential execution.

Google (Gemini) Parameters

Parameter	Range	Default	Description
Thinking Level	LOW, HIGH	HIGH	Controls depth of internal reasoning. LOW = faster responses, HIGH = deeper analysis. Only for Gemini 2.5 and 3 models.
Temperature	0–2	1	Controls output randomness.
Top P	0–1	0.95	Nucleus sampling. Note: Google’s default (0.95) is lower than other providers.
Top K	1+	Auto	Maximum tokens considered at each generation step.
Max Output Tokens	1+	Auto	Maximum tokens to generate.

Temperature vs Reasoning: Temperature controls randomness in word selection. Reasoning effort/thinking controls how much the model deliberates before responding. For complex analysis, increase reasoning, not temperature.

Manages long conversations by summarizing older messages when context limits are reached.

​

How It Works

When a conversation approaches the model’s context limit, the system:

1. Identifies the oldest messages that can be compacted
2. Generates a structured summary preserving key information
3. Replaces compacted messages with the summary
4. Keeps recent messages in full detail

The summary follows a structured format: Goal, Actions Taken, Key Data, Status, and Next Steps (approximately 300 words). This preserves actionable context while reducing token usage.

Compaction operates at the part level within messages, not just message level. This enables finer-grained compaction when a single message contains many large parts (e.g., 24 tool call results totaling 150k+ tokens).

​

Auto vs Override Mode

Behavior	Auto Mode (Default)	Override Mode
When to summarize	System determines automatically	System determines automatically
Summary model	Auto-selected (fast, cost-effective)	You choose the model
Prune Protection	Optimized default (40,000 tokens)	Configurable
Summary Max Tokens	Optimized default (30,000 tokens)	Configurable

Override Parameters:

Setting	Default	What It Controls
Summary Model	Auto	Which model generates summaries. Auto picks a fast, cost-effective model.
Prune Protect Tokens	40,000	Recent tokens always kept in full detail (never summarized).
Summary Max Tokens	30,000	Conversation length threshold before summarization triggers.

Keep defaults unless you have specific context requirements. Lower Prune Protect = more aggressive summarization. Higher Summary Max = summaries trigger earlier.

Ensures reliability by automatically switching to alternative AI models when your primary model is unavailable.

​

How It Works

When errors occur, the system:

1. Classifies the error type (rate limit, provider error, network, timeout)
2. Retries based on error severity
3. Falls back to the next model in the chain if retries are exhausted
4. Automatically excludes models from the same provider (ensures true redundancy)

Retry Behavior by Error Type:

Error Type	Retries Before Fallback	Rationale
Rate Limit	2	Rate limits often clear quickly
Provider 5xx	1	Provider unhealthy, fallback faster
Network Error	0	Immediate fallback (no retries)
Timeout	1	Server overloaded, try once then fallback

​

Auto vs Override Mode

Auto Mode (Default): The system selects fallback models based on your primary model’s characteristics:

Primary Model Type	Fallback Chain
Expert/Smartest (intelligence rating > 4)	Claude Opus 4.5 → Gemini 3 Pro → GPT-5.2
Fastest (speed rating > 4)	Gemini 3 Flash → Claude Haiku 4.5 → GPT-4.1
Recommended/Balanced	Claude Sonnet 4.5 → Gemini 3 Flash → GPT-5.2

Override Mode: Manually select up to 2 fallback models. Drag-and-drop to set priority order. Models from the same provider as your primary are automatically filtered out.

Disabling fallback means your agent will fail if the primary model is unavailable. Keep enabled unless you have specific compliance requirements.

​

Understanding Credit Costs

​

How Credits Work

• Message length: Longer messages consume more credits
• Model selected: Different AI models have different credit costs
• Conversation history: Each message includes previous context
• Tools available: More tools slightly increase base cost

​

Model Pricing Overview

​

Workflow and Integration Costs

• Free: Google Workspace, Slack, Airtable, Notion, Salesforce, HubSpot, 150+ more
• Paid: Data enrichment services (eg. Hunter.io ~10 credits, ZoomInfo ~60 credits)

​

Optimizing Credit Usage

​

Tracking Credit Usage

• Real-time Display: Credits shown next to each agent response as it streams
• Conversation History: View total credits per conversation thread
• History Page: See credit costs for agent chats and Agent node executions (when run via workflows)
• Account Dashboard: Track total credits used across all agents and timeframes

​

Credentials & Authentication

​

How Agents Use Credentials

​

Data Privacy & Security

​

Best Practices

Never:
- Delete customer data or records
- Send emails without explicit user approval
- Make purchases or financial commitments
- Override manual decisions by team members
- Modify production data without confirmation

​

Troubleshooting

​

Next Steps