LetTokenBurnRealCLI: Toward a Unified Evaluation of Human-Agent Collaboration for CLI Agents
A paired benchmark with 101 CLI tasks evaluating task execution and 5-dimensional human-agent collaboration quality via autonomous and scripted interactive modes.
$ npm i -g lettokenburn-cli$ ltb bench info realcli
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| Claude Opus 5 | 49.2 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| DeepSeek V4.1 Flash | 45.6 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Qwen3.8 Max | 42.7 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Claude Opus 4.8 | 40.8 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Kimi K3 | 38.4 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| DeepSeek V4 Flash | 37.0 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| GPT-5.6 Sol | 35.7 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| GLM-5.3 | 33.9 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| GLM-5.3 Flash | 33.4 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| Hy3 | 22.7 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
| MiniMax-M3 | 22.7 | N/A | Independent | medium | Oct 15, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
RealCLI comprises 101 realistic CLI tasks across 10 categories, evaluated in two modes: EXEC (full specs) and COLLAB (scripted progressive user interaction). It measures task outcomes and collaboration via a five-dimensional framework including steerability and safety.
Scoring uses environment verifiers for task success and trace-based checklists for collaboration. Frontier agents achieve low scores (peak joint collaboration score 37.6%), revealing significant gaps in trust calibration and safety despite moderate task success.