Tura agent cuts LLM token use 80%, lifts success rate to 80%

Tura is a local, open-source coding agent designed for developers skeptical of agent claims. The announcement centers on benchmark results from 60 sessions across 20 DeepSWE v1.1 tasks, using GPT-5.6 SOL at High reasoning effort. Tura offers two usage modes: Balanced and Direct. Balanced prioritizes success rate and reasoning quality, achieving an 80.0% success rate while using 49.6% fewer tokens than the official Codex CLI High configuration. Direct prioritizes cost, cutting aggregate tokens by 83.5% with a 65.0% success rate versus Codex CLI High's 60.0%. Both outperform Codex CLI High's baseline.

The technical approach rests on three main systems. First, a macro CLI command tool consolidates multi-step workflows. A traditional tool-calling agent might require five separate LLM turns to inspect code, apply a patch, build, run tests, and lint. Tura bundles these into a single command_run structure with sequential steps (shell_command, apply_patch, etc.), allowing the model to design and execute a multi-step sequence in one turn. In the benchmark, Tura Balanced achieved 66.8% fewer model rounds and 49.6% fewer tokens than Codex CLI High; Direct achieved 84.0% fewer rounds and 83.5% fewer tokens.

Second, backward reasoning guides the model away from statistically common but weak solutions. Rather than reasoning forward from current state to goal (s₁ → s₂ → ... → sₙ), Tura prompts the model to reason backward, estimating the desired end state first (sₙ₋₁) and then working backward to root causes and optimal logic. The example given is rock-paper-scissors: a standard forward prompt might produce a biased result (since LLMs encode text-token probability distributions), but backward reasoning allows the model to recognize that true fairness requires external randomness and to generate correct pseudocode using randint(1, 3). In coding tasks, this translates to reconstructing failure states and identifying root causes before writing code. Tura Balanced passed 12 more of 60 binary verifiers than Codex CLI High—a 20-point success gap. The article notes this is a system-level association, not a causal proof, and both systems used GPT-5.6 SOL at High reasoning, ruling out effort-level mismatch as an explanation.

Third, runtime context management prevents the token bloat that accumulates in traditional skill-based agents. Instead of loading broad Markdown skills into a long-lived session and letting them pile up until compaction, Tura ties runtime prompts to explicit task state. Sessions can be renamed, refreshed, and managed via CLI; task-specific manuals and commands load through a recursive task tree; and irrelevant context is removed or compacted. Checkpoints retain code locations, patches, tests, and task status rather than loose summaries. In archived benchmark sessions, Tura resumed execution an average of 2.6 rounds after compaction, versus an estimated 5.4 rounds for Codex (estimated from token-count drops, since Codex does not expose compaction events).

Tura is available via npm (npm install -g tura-ai) for Mac, Linux, and Windows, or via source checkout from GitHub. It supports multiple entry points: a terminal UI (tura), a direct CLI runner (tura exec "prompt"), a gateway-backed runner (tura run "prompt"), and a desktop GUI (tura_gui). On first launch, users must configure an LLM provider and select a model; Tura does not bundle credentials. The documentation is published on GitBook-style docs/SUMMARY.md and covers installation, CLI parameters, task status, context management, runtime prompts, command structure, agents, personas, and custom providers. The project is open-source under AGPL and includes a contribution guide emphasizing small, evidence-backed changes and test ownership.

The published benchmark has documented limits. Testing covers only GPT-5.6 SOL at High reasoning on 20 DeepSWE tasks and 5 rewrite tasks (with 2 design tasks kept outside the scored population). Broader testing with Anthropic Claude, Google Gemini, OpenAI-compatible models, local providers, UI latency, cross-OS measurements, and runtime/session parsing remain on the roadmap and are listed as known evidence gaps.

Tura addresses a core pain point in AI agent design: the gap between polished benchmarks and real-world developer experience. Traditional coding agents rely on repetitive tool-calling loops—inspect, patch, build, test, lint—that accumulate conversational overhead with every step. By consolidating these into a single macro command, Tura eliminates the statistical burden of choosing the next action and reduces token waste on model round trips that serve only to maintain context.

The backward-reasoning strategy reveals a deeper insight: LLMs, as statistical models, naturally follow high-probability patterns in training data, which often mean common but under-considered code. Tura decouples this statistical bias by asking the model to work backward from the goal state—first imagining what a correct solution should look like, then reconstructing the failure scenario and root cause. This shift from forward chaining (current state → goal) to backward chaining (goal → intermediate states → root cause) mirrors classical planning and appears to guide the model away from mediocre defaults.

The benchmark results are system-level associations (Balanced vs. Codex CLI High shows a 20-point success lift while using half the tokens), not isolated causal proofs for any single feature. The published report acknowledges this—comparing only GPT-5.6 SOL at the High reasoning level across DeepSWE and rewrite tasks. Broader testing with other models and providers remains a documented gap. Context management through task_status and runtime prompts also plays a role: Tura reports resuming from compaction checkpoints after an average of 2.6 rounds, versus an estimated 5.4 for Codex, suggesting that precise runtime state preservation prevents the stale-context drift that plagues long-lived sessions.

No comments yet. Be the first to share your thoughts!