Cognosys

Cognosys breaks down user-provided goals into discrete subtasks using an LLM-based planning loop, then executes each subtask sequentially with feedback loops. The system maintains execution state across steps, allowing it to recover from failures and adapt subsequent tasks based on prior results. This implements a goal-oriented agent architecture similar to AutoGPT's task queue pattern, where each step is evaluated before proceeding to the next.

Cognosys provides a schema-based function registry that maps user intents to external APIs and web services (search engines, data APIs, automation platforms). The system uses function-calling patterns to invoke these tools within the task execution loop, parsing responses and feeding results back into the planning context. This enables the agent to interact with external systems without requiring users to write integration code.

Cognosys allows users to customize the system prompts and reasoning patterns used by agents through a visual prompt editor. Users can define agent personality, reasoning style, constraints, and output format without modifying code. The system supports prompt templates with variable substitution, few-shot examples, and chain-of-thought instructions. Changes to prompts are immediately reflected in subsequent task executions, enabling rapid iteration on agent behavior.

Cognosys renders a live task execution tree in the browser, displaying each subtask's status (pending, running, completed, failed) with streaming output from the LLM. Users can pause execution, inspect intermediate results, manually override task parameters, or inject new instructions mid-execution. This is implemented via WebSocket connections to the backend that push execution state updates in real-time, allowing synchronous human-in-the-loop control.

Cognosys accepts free-form natural language descriptions of goals and uses an LLM to translate them into structured task plans with estimated execution time, resource requirements, and success criteria. The system infers task dependencies, identifies required tools, and generates subtask descriptions without user intervention. This leverages prompt engineering and few-shot examples to map user intent to executable task graphs.

Cognosys maintains execution context across task steps by storing intermediate results, tool outputs, and LLM reasoning in a context window that is passed to each subsequent task. The system implements a sliding window approach to manage token limits, prioritizing recent results and user-specified critical information. This enables tasks to reference prior results without explicit data passing, simulating a working memory for the agent.

When a task fails (API error, timeout, invalid output), Cognosys automatically analyzes the error, generates a corrected task variant, and retries with modified parameters or alternative tools. The system uses LLM-based error diagnosis to determine if the failure is transient (retry with backoff) or structural (modify approach), and implements exponential backoff with jitter for transient failures. Failed tasks can be manually re-executed with user-provided corrections.

Cognosys integrates web search APIs (Google, Bing, or similar) as a built-in tool that agents can invoke to fetch real-time information. The system automatically parses search results, extracts relevant snippets, and feeds them into the task context. Search queries are generated by the LLM based on task requirements, and results are ranked by relevance before inclusion in context. This enables agents to access current information beyond their training data cutoff.

Cognosys records all task executions with full audit trails including task parameters, LLM prompts, tool invocations, results, and timestamps. Execution logs are stored in a queryable format and can be exported as JSON, CSV, or PDF reports. The system provides a history browser UI where users can replay past executions, compare different runs, and identify patterns or failure modes. Logs include sensitive data masking options for compliance.

Cognosys allows users to schedule task executions on a recurring basis (hourly, daily, weekly, monthly) using cron-like syntax or a visual scheduler. Scheduled tasks run on Cognosys infrastructure without requiring the user's browser to be open. The system manages execution state, handles missed runs, and provides notifications on completion or failure. Scheduling integrates with the task decomposition and monitoring capabilities, allowing users to observe scheduled executions in real-time.

Cognosys supports creating multiple autonomous agents that can collaborate on complex tasks by delegating subtasks to each other. Agents communicate via a message-passing system where one agent can invoke another agent as a tool, passing task parameters and receiving results. This enables hierarchical task decomposition where high-level agents break down goals and delegate specialized subtasks to domain-specific agents. Agent state and results are shared through a common context store.