Otter Bot is an AI-driven agent orchestration platform designed to automate complex, multi-step workflows by allowing users to delegate tasks to autonomous "bot teams." Unlike standard chat-based AI interfaces that simply provide information, Otter Bot executes real-world actions, such as file manipulation, software execution, and browser interaction, to complete tasks from start to finish.

Core Methodology and Technical Architecture

1. Agentic Execution Environment (Bot Computers):

Each bot operates within a sandboxed, isolated Linux desktop environment managed via Docker. Each containerized instance is assigned its own dedicated terminal, browser, and file system. This architectural separation ensures that bot sessions, login cookies, and local browser data do not conflict with the user's primary environment. To maintain security, these containers are accessible only via 127.0.0.1 using randomized VNC passwords and do not mount the host file system directly.

2. Workflow Orchestration:

The system employs a hybrid execution model where tasks are organized into groups. It manages both sequential dependencies—where $\text{Task}_{n+1}$ initiates only upon the successful completion of $\text{Task}_n$—and parallel execution, where sub-agents perform independent actions simultaneously to optimize throughput. Users track the state of these processes via a centralized "Plan Timeline," providing transparency into the agentic decision-making process.

3. Computer Use and Vision:

The platform utilizes screen-reading capabilities to interact with legacy web interfaces or software lacking APIs. The agent interprets the desktop GUI by analyzing screenshots, identifying interactive elements (assigned numeric indices), and issuing commands such as browser_click or browser_type. All interactions are logged, creating a verifiable audit trail of tool usage.

4. Integrations and BYOK (Bring Your Own Key):

Otter Bot supports a flexible model-agnostic layer. It allows users to leverage existing subscriptions (e.g., Anthropic’s Claude Code, OpenAI’s Codex) via OAuth, or integrate custom endpoints through API keys or local GPU servers compliant with OpenAI’s protocols. This enables the deployment of specialized models (like DeepSeek or local models) tailored to the specific nature of the agent's task.

5. Safety and Human-in-the-Loop (HITL) Controls:

The platform implements a tiered security model:

• Scope Limitation: Local tools are restricted by realpath to operate strictly within the defined working folder, preventing access to sensitive directories like ~/.ssh.
• Permissioning: Sensitive operations, such as system commands, email dispatch, or financial transactions, trigger an explicit approval card. Users must interact with these "Commit/Cancel" prompts to authorize execution.
• Isolation: The UI bridge uses contextIsolation and standard web-sandbox technologies to prevent malicious scripts from interacting with the host OS.

6. Routine Management:

The system includes a scheduler for recurring tasks, enabling agents to monitor metrics (e.g., invoice retrieval, sales forecasting) at defined temporal intervals (e.g., $T = \text{Monday } 09:00$). Mobile parity ensures that human oversight—specifically the authorization of high-stakes actions—remains possible even when the user is away from the desktop environment.