2. Agent Management

Agents are the heart of your Sutra organization. Each agent is a specialized AI worker with its own personality, tools, and expertise — and how you configure them determines how effectively your entire autonomous organization operates. This guide covers everything from creating your first agent to building full team hierarchies.

Managing Agents

The Agents dashboard (/agents) provides a comprehensive list of all AI personas available in your system. From here, you can create, clone, edit, or archive agents.

Each agent card shows its name, avatar, active status, assigned purpose, and tool count at a glance. You can filter by team, status, or model provider to quickly find what you need.

Anatomy of an Agent

When you create or edit an agent, you configure several critical dimensions:

1. Identity: Name, Avatar, and a short Description. This helps human users identify the agent's role at a glance.
2. System Prompt: The core instructions that define the agent's personality, formatting rules, and operational boundaries. A well-crafted system prompt is the single biggest factor in agent quality. Be specific about the role, the expected output format, and any constraints. For example, a code-review agent might include: "You are a senior code reviewer. Always check for security vulnerabilities, performance issues, and adherence to the team's TypeScript style guide. Provide feedback as inline comments with severity labels (critical, warning, suggestion)."
3. LLM Configuration:
4. Direct Model Selection: You can force an agent to use a specific provider (e.g., Anthropic) and model (e.g., claude-3-5-sonnet).
5. Smart Routing (Purpose) [Recommended]: Instead of locking an agent to a specific model, you assign it a "Purpose" (e.g., "Heavy Reasoning"). Sutra's Smart Router dynamically selects the best available model based on real-time rate limits, cost, and provider availability. This means if OpenAI is experiencing an outage, your agents automatically fall back to Anthropic or a local Ollama model — zero downtime, no manual intervention. You configure the priority waterfall in the System & Configuration section.
6. Behavioral Parameters:
7. Temperature: Controls creativity vs. determinism (0.0 for strict coding, 0.7 for creative writing).
8. Max Tokens: Limits the length of the agent's response.
9. Tool Arsenal: Toggle the specific tools this agent is allowed to use (e.g., search_knowledge_base, read_file, run_shell_command).

Notable Tool Categories

Sutra ships with 50+ built-in tools organized by category. Here are some of the most powerful ones:

• Browser Automation: A full interactive browser toolkit (browser_open, browser_click, browser_type, browser_screenshot, browser_extract_text, browser_extract_data, browser_scroll, browser_navigate, browser_select, browser_wait). Agents can navigate websites, fill forms, take screenshots, and extract structured data — all with SSRF protection that blocks navigation to localhost and private IP ranges. Sessions persist across tool calls within a conversation, so agents can complete multi-step web workflows. Sutra also supports session recording (browser_record_start, browser_record_stop) to capture browser automation sequences.
• Data Analysis: The analyze_data tool lets agents load CSV or Excel files into a pandas DataFrame and execute Python code to analyze them. The code runs in a sandboxed environment — import statements, eval, exec, and dunder attribute access are all blocked. Only pandas and the loaded DataFrame are available, keeping agent-submitted code safe.
• Playbooks: Pre-written browser automation scripts stored as Markdown files with YAML frontmatter. Agents can list_playbooks and load_playbook to execute repeatable browser workflows (e.g., "check deployment status on Vercel" or "scrape competitor pricing").
• Developer Tools: GitHub (commit_and_push, create_github_pr, create_github_issue), GitLab, Jira, Linear, and Notion integrations for project management and code collaboration.
• OS & File System: read_file, write_file, list_directory, search_files, run_shell_command, get_system_info, send_notification. File access is restricted to explicitly configured allowed paths.
• Communication: send_email (SMTP), send_telegram_message, post_to_slack_channel, and webhook delivery tools.

Skills: Reusable Capabilities

As your system grows, you will find that multiple agents need similar instructions (e.g., "How to write idiomatic React code" or "How to format a daily report"). Instead of copy-pasting these into every system prompt, Sutra uses Skills (/skills).

How Skills Work

A Skill is a modular block of knowledge or instructions. - Dynamic Injection: When a Skill is attached to an agent, its content is dynamically injected into the agent's context window. - Variables: Skills can include variables (e.g., {{user_name}} or {{current_date}}) that are resolved at runtime. - Centralized Updates: If your company coding standards change, you update the "Coding Standard" Skill once, and every agent using that skill instantly inherits the new behavior.

Templates: Rapid Provisioning

Templates (/templates) allow you to bootstrap new agents instantly using community-proven or organization-approved configurations.

Using Templates

When spinning up a new project, you might need a standard "Software Engineer" agent. By selecting the Software Engineer template, Sutra automatically populates: - The optimal System Prompt. - The recommended LLM configuration. - The necessary tools (e.g., File System access, Git integrations).

You can save any of your existing, well-performing agents as a custom Template for your organization to reuse.

Built-in Templates

Sutra ships with 11 pre-built templates covering common organizational roles:

• CEO Agent (leadership) — Sets organizational strategy, delegates to department leads, and approves major decisions.
• Product Manager (management) — Breaks strategic goals into actionable tasks, manages roadmap, and coordinates across teams.
• Software Engineer (engineering) — Implements features, fixes bugs, writes code, and handles technical execution.
• Marketing Specialist (marketing) — Creates content, runs campaigns, manages brand communications, and tracks performance.
• Finance Analyst (finance) — Tracks costs, manages budgets, produces financial reports, and flags anomalies.
• Research Specialist (research) — Conducts deep research, synthesizes information, and produces comprehensive reports.
• Customer Success Agent (operations) — Handles customer interactions, resolves issues, and drives satisfaction.
• Security Auditor (security) — Audits systems for vulnerabilities, enforces security policies, and responds to incidents.
• Data Analyst (data) — Analyzes datasets, builds reports, identifies trends, and generates actionable insights.
• HR Manager (management) — Manages agent onboarding, performance reviews, and organizational health.
• General Assistant (general) — Versatile assistant for general tasks, Q&A, and ad-hoc automation.

Each template includes a detailed system prompt, recommended tools, and LLM configuration tuned for that role. Templates are the fastest way to staff your autonomous organization — select a template, customize the name and avatar, and you have a production-ready agent.

Agent Check-ins

Running agents can perform scheduled self-assessments called check-ins. By default, check-ins run daily at 8:00 AM Pacific (configurable via the checkin_cron setting).

During a check-in, the agent reviews its assigned goals and tasks, then produces a structured report containing:

• Summary: A brief overview of the agent's current state.
• Goals reviewed: Progress updates on each assigned goal.
• Tasks reviewed: Status of assigned tasks with notes.
• Blockers: Anything preventing progress.
• Proposed actions: Next steps the agent plans to take.
• Stuck items: Tasks or goals flagged as stuck (no progress for too long).
• Proposed initiatives: New project or task ideas the agent suggests.

Check-ins make it easy to get a daily pulse on your entire autonomous organization without manually querying each agent.

Organization Chart & Hierarchy

Sutra allows you to construct explicit reporting structures using the Organization Chart (/org).

Roles and Teams

• Roles: Define the job title (e.g., "Lead Reviewer", "Data Scraper").
• Teams: Group agents logically (e.g., "Frontend Team", "Marketing Team").

Reporting Lines

By configuring who reports to whom, you enable autonomous delegation. - If the "Lead Reviewer" agent is assigned a massive codebase refactor, it can autonomously break the task down and assign sub-tasks to the "Junior Coder" agents that report to it. - Once the "Junior Coder" finishes, the system routes the result back up the hierarchy for the "Lead Reviewer" to validate. This hierarchical collaboration mirrors human corporate structures and dramatically increases the quality of the final output.