Configuring Agents

Every agent in Pencel has a set of configuration options that control how it behaves, how much autonomy it has, and which AI model powers it. Getting these settings right makes the difference between an agent you trust and one you are constantly babysitting.

TIP

Two often-overlooked fields: Description (a short summary shown in agent lists — also used to identify the right agent in multi-agent workflows) and Status (active vs inactive — deactivate an agent to temporarily remove it from schedules without deleting its configuration).

Risk tolerance

Risk tolerance controls how cautious or aggressive an agent is when making decisions and taking actions.

Level	Best for	Examples
Low	Sensitive data, financial decisions, customer-facing content	Processing invoices, drafting compliance reports, sending client emails
Medium	Routine business tasks with moderate stakes	Weekly summaries, internal reports, data organization
High	Exploration, research, and rough drafts	Brainstorming ideas, competitive research, internal memos

TIP

Start with low risk tolerance for any new agent. You can always increase it later once you see how the agent performs and build confidence in its judgment.

• Low means the agent asks for confirmation more often and avoids actions with uncertain outcomes. Use this whenever the output goes to customers or touches financial data.
• Medium is the right default for most internal work. The agent takes reasonable actions on its own but pauses for anything unusual.
• High gives the agent more freedom to experiment. Great for research and brainstorming, but not recommended for anything that goes outside your team.

Writing style

The writing style setting shapes the tone, format, and voice of everything the agent produces. Write it in plain language, as if you are describing how you want a colleague to write.

Good writing style instructions:

• "Professional and concise. Use bullet points over paragraphs. Avoid jargon."
• "Friendly but authoritative. Write as if explaining to a smart non-expert."
• "Executive-level. Lead with conclusions, support with data, keep it under 200 words."

INFO

Writing style applies to all of the agent's output — chat responses, workflow results, and artifacts. If you need different styles for different tasks, create separate agents.

Model selection

Pencel supports 7 models across 3 providers (Anthropic, Google, OpenAI). See Supported Models for the full list with pricing details.

How to choose:

• Default to Anthropic Claude for any task that involves judgment, complex reasoning, or important output. It produces the most reliable results for business-critical work.
• Use Gemini Flash or GPT-4o mini for high-volume, routine tasks like formatting data, simple extraction, or categorization. You save significantly on costs without sacrificing quality for straightforward work.
• Use GPT-4o or Gemini Pro when you need a balance of capability and cost, or when you want to compare results across providers.

TIP

You can change an agent's model at any time. A common pattern is to prototype with a fast, cheap model and then switch to Claude for production runs.

Plan approval

This setting controls whether the agent shows you its plan before executing it.

• Auto-approve: The agent creates a plan and immediately starts executing. Best for trusted, well-tested workflows where you are confident the agent will do the right thing.
• Require approval: The agent creates a plan and waits for your review before proceeding. You see each planned step and can approve, modify, or reject the plan.

WARNING

Use require approval for any agent that modifies external systems (sends emails, updates databases, posts to Slack). Even a well-configured agent can misinterpret instructions — the approval step gives you a safety net.

When to use each:

Scenario	Recommended setting
Agent is new and untested	Require approval
Workflow touches customer data	Require approval
Routine internal report you have run 20 times	Auto-approve
Research and brainstorming tasks	Auto-approve
Anything that sends data outside your organization	Require approval

Set a cost cap

The Max Cost Per Run setting stops an agent automatically if a single run exceeds the dollar limit. This prevents runaway costs from loops or unexpectedly long tasks.

Agent role	Suggested cap
Simple formatter or extractor	$0.25
Report writer	$1.00
Research analyst	$3.00
General-purpose (default)	$5.00

TIP

Always set a cost cap — even a generous one. Without it, an agent stuck in a reasoning loop can spend indefinitely. You can adjust after reviewing real costs on the Jobs page.

Restrictions

Restrictions are behavioral constraints specific to one agent — unlike guidelines, which apply across your workspace.

Use restrictions for rules that only matter for a particular agent's role:

• "Never send emails to external addresses."
• "Do not modify or delete existing artifacts — only create new ones."
• "Always ask for approval before querying the production database."

When to use which: If every agent should follow the rule, make it a guideline. If only this agent needs it, use a restriction.

Output length and citations

Setting	Options	When to change
Output length	concise / standard / detailed	Use concise for data extraction agents, detailed for report writers
Citation policy	always / when_available / never	Use always for compliance or audit-facing agents that must trace every claim to a source

These defaults (standard and when_available) work well for most agents. Adjust when the agent consistently over- or under-delivers on length, or when citations are a firm requirement.

Enabled skills

Enabled skills determine which tools and capabilities the agent can use. By default, agents have access to all built-in tools. You can limit this for safety or focus.

Why limit skills:

• Safety. An agent that only needs to search and summarize should not have the ability to create or modify artifacts.
• Focus. Fewer available tools means the agent spends less time deciding which tool to use and more time doing the actual work.
• Cost. Tool definitions are included in every API call. Fewer tools means fewer input tokens.

Common configurations:

Agent role	Recommended skills
Report writer	Search workspace, create artifacts
Data analyst	Search workspace, list entities, get costs
Operations manager	All built-in tools + relevant integrations
Read-only reviewer	Search workspace, list entities

INFO

Enabled skills only affect what the agent can do. The agent still follows its workflow instructions — it will not randomly use tools just because they are available.

Putting it all together

Here is an example configuration for a "Weekly Report Writer" agent:

Setting	Value	Why
Risk tolerance	Medium	Internal reports, moderate stakes
Writing style	"Professional, concise. Use tables and bullet points. Lead with key findings."	Matches leadership team preferences
Model	Anthropic Claude	Report quality matters more than speed
Plan approval	Auto-approve	This workflow has been tested and refined
Max Cost Per Run	$1.00	Weekly reports rarely exceed $0.30 — this catches anomalies
Output length	Standard	Reports should be thorough but not exhaustive
Enabled skills	Search workspace, create artifacts	Only needs to read data and produce output
Restriction	"Do not include individual employee names in reports."	Privacy requirement for leadership summaries

Start with conservative settings and loosen them as you build confidence. It is always easier to give an agent more freedom than to undo something it should not have done.