Oleksandr Kuzmenko, EditorPublished 6 min readAI-assisted · editor-reviewed
TL;DR
- Agav halves token expenditures by isolating subagent context and bounding error loops.
- Terminal-Bench 2.1 trials demonstrate a 49.1% cost drop against a 16.2 percentage-point drop in accuracy versus Codex.
- The client runs as a standalone binary with native sandboxing, encrypted keys, and CI-ready JSON Schema output.
Key facts
- Cost savings vs Codex
- 49.1% lower total spend (Terminal-Bench 2.1)
- Median trial cost
- 15.2¢ across 89 benchmark tasks
- SWE-bench-Live Lite verification
- 62.0% (186/300 tasks)
- Terminal-Bench 2.1 accuracy
- 84.7% (377/445 trials, maintainer-reported)
- Open source license
- Apache 2.0
Harness Optimization Over Model Scaling
AI coding bills escalate rapidly when terminal agents re-send unbroken conversation logs and massive tree listings. Agav mitigates this by treating cost reduction as an engineering constraint in the agent harness. It leverages prompt-cache reuse, bounds retry sequences, breaks recursive error loops, and delegates exploratory commands to isolated subagents with fresh context.
Benchmarks and Trade-Offs
According to data published by the maintainers, Agav underwent cost evaluations on Terminal-Bench 2.1 against Codex using GPT-5.6 Luna at identical pricing rates:
- Median Trial Cost: 15.2¢ across 89 tasks.
- Cost Reduction: 49.1% lower total expenditure than Codex.
- Accuracy Compromise: Agav achieved a 60.4% pass rate compared to 76.6% for Codex.
- SWE-bench-Live Lite: 62.0% verification rate (186/300 tasks).
- Terminal-Bench 2.1 Accuracy: 84.7% (377/445 trials).
Runtime Architecture and Tooling
Agav ships as a single compiled executable with zero Node.js or Python runtime requirements. It features native OS sandboxing on Linux and macOS, fine-grained per-action permissions, encrypted credential storage, and an interactive /compact command to reclaim context without wiping execution plans. Non-interactive CLI runs can be integrated directly into CI via agav run with structured JSON Schema output.
Try it in 2 minutes
curl -fsSL https://agav.dev/install.sh | bash
agav --provider openrouter --model openrouter/free run "List files and explain architecture"When to use
- Use when running autonomous agents over large repositories with strict API token budgets.
- Use for CI automation pipelines where commands require OS sandboxing and JSON Schema validation.
When NOT to use
- Do not use when maximum task completion accuracy is paramount over API spend, where Codex currently scores 16 percentage points higher.
- Avoid using in enterprise production pipelines that require a verified, long-term stable API, as Agav is currently in experimental beta.
What to do today
- Install Agav via curl or release binaries to test locally with free OpenRouter or NVIDIA NIM credits.
- Use /compact during long CLI debugging sessions to reclaim context tokens while retaining execution plans.
- Configure strict per-action sandbox permissions when running automated commands in agav run.
Step names are taken from What to do today.
Step names are taken from What to do today. Install Agav via curl or release binaries to test locally with free OpenRouter or NVIDIA NIM credits.. Use /compact during long CLI debugging sessions to reclaim context tokens while retaining execution plans.. Configure strict per-action sandbox permissions when running automated commands in agav run.
- Install Agav via curl or release binaries to test locally with free OpenRouter or NVIDIA NIM credits.
- Use /compact during long CLI debugging sessions to reclaim context tokens while retaining execution plans.
- Configure strict per-action sandbox permissions when running automated commands in agav run.
What the community says
at agav we focus on two main verticals which are cost efficiency with minimal accuracy tradeoffs and secure execution through permissions, approvals and OS native sandboxing.
What practice do you use to minimise the token use?
