Omar – A TUI for managing 100 coding agents

Manages spawning, monitoring, and termination of up to 100 concurrent coding agents through a centralized TUI control plane. Uses event-driven architecture to track agent state (idle, running, completed, failed) and coordinates resource allocation across agents. Implements agent pooling with configurable concurrency limits to prevent resource exhaustion while maintaining responsiveness of the terminal interface.

Renders a live-updating TUI dashboard displaying individual agent states, progress indicators, resource usage (CPU, memory), and execution logs. Uses asynchronous event streams from agents to update display components without blocking user input. Implements efficient terminal rendering with dirty-region updates to minimize flicker and reduce CPU overhead.

Saves agent state, task queue, and execution progress to disk, enabling recovery from crashes or intentional shutdowns. Implements checkpoint-based recovery where agents can resume from the last successful checkpoint. Supports exporting and importing sessions for reproducibility and sharing.

Captures detailed logs from all agents (LLM API calls, intermediate reasoning, errors) and provides structured access to logs for debugging. Implements log filtering, search, and export capabilities. Supports multiple log levels (debug, info, warn, error) configurable per agent. Integrates with standard logging frameworks (syslog, JSON logging).

Distributes incoming coding tasks across available agents using configurable scheduling strategies (round-robin, least-loaded, priority-based). Implements backpressure handling to queue tasks when all agents are busy, with optional timeout and retry logic. Tracks task dependencies and ensures agents receive tasks in correct order if sequential execution is required.

Allows users to define agent profiles with specific capabilities, model backends (GPT-4, Claude, local LLM), and behavioral parameters (temperature, max tokens, system prompts). Stores configurations in a declarative format (YAML or JSON) and validates them at startup. Enables dynamic agent spawning based on configuration templates, allowing rapid scaling without code changes.

Collects outputs from all agents (code, logs, metrics, errors) and aggregates them into a unified result set. Implements deduplication logic to remove duplicate solutions from multiple agents, and ranking/filtering to surface highest-quality results. Supports multiple output formats (JSON, CSV, structured code diffs) for downstream processing.

Provides TUI-based controls to pause, resume, kill, or restart individual agents or groups of agents without stopping the entire system. Implements interactive prompts for agent-specific actions (e.g., modify system prompt mid-execution, adjust temperature). Supports keyboard shortcuts and mouse interactions for rapid control without context switching.

Monitors agent health through heartbeat signals and timeout detection, automatically identifying crashed or hung agents. Implements configurable recovery strategies (restart, reassign tasks, escalate to user). Tracks failure history per agent and can disable agents that exceed failure thresholds to prevent cascading failures.

Collects and aggregates performance metrics from all agents (execution time, token usage, success rate, output quality). Provides analytics views showing agent efficiency, throughput, and comparative performance. Exports metrics in standard formats (Prometheus, JSON) for integration with external monitoring systems.

Allows users to organize agents into logical groups (by capability, by model, by project) and perform batch operations on groups (pause all agents in group, update config for group, collect results by group). Implements group-level metrics and status aggregation. Supports hierarchical grouping for complex fleet structures.

Enables agents to share results, request help from other agents, or coordinate on complex tasks. Implements a message passing system with optional message queuing and filtering. Supports both direct agent-to-agent communication and broadcast messages to all agents. Includes conflict resolution for competing agent outputs.