Tura-AI/tura

What it solves

Tura addresses the inefficiency of traditional ReAct-style agent loops, where models must repeatedly re-enter the session after every tool call. This cycle often leads to high token consumption due to the repeated transmission of system prompts and growing context, and can result in mediocre logic because models tend to follow statistically common but suboptimal paths.

How it works

Tura employs three primary strategies to optimize agent performance:

• Macro CLI Commands: Instead of multiple small tool calls, Tura uses a command_run macro tool. This allows an agent to build a multi-step execution tree (e.g., inspect, patch, build, test) and execute it in a single LLM turn, drastically reducing round trips.
• Backward Reasoning: Tura guides the LLM to estimate the goal state ($s_n$) and reason backward to the current state, rather than reasoning forward. This helps the agent identify root causes and derive optimal strategies before writing code.
• Runtime Context Management: Tura treats context as a state machine. It uses task_status, runtime prompts, and recursive execution manuals to keep context scoped to the current task, allowing for precise compaction of irrelevant data without losing critical execution details.

Who it’s for

Tura is designed for developers building long-horizon autonomous agents, particularly those focused on complex coding tasks that require multi-step verification and high token efficiency.

Highlights

• Token Efficiency: Reduces aggregate token usage by up to 77.5% compared to Codex CLI in specific configurations.
• Improved Success Rates: Demonstrated a 16.7 percentage point higher success rate than Codex CLI on DeepSWE tasks in "Balanced" mode.
• Flexible Interfaces: Provides a CLI, TUI, and a desktop GUI workspace client.
• Extensible Architecture: Supports custom providers, personas, agents, and commands.