Research / Coding Agents
The model is only part of the agent.
A coding agent’s performance depends on more than the model it uses. It also depends on how it manages context, selects tools, delegates work, and handles failure.
R-CLI is Backboard’s terminal coding agent and a platform for investigating those decisions.
Same model. Different results.
We evaluated R-CLI with Claude Opus 4.8 on Terminal-Bench 2.1, a benchmark of tasks that require agents to work in a terminal.
Configuration
Task completion
Backboard R-CLI + Claude Opus 4.8
84.04%
Claude Code + Claude Opus 4.8¹
78.9%
374 successful trials. 89 tasks. Five attempts per task.
R-CLI scored 5.14 percentage points above the published same-model comparison recorded in our report. It completed 65 of the 89 tasks successfully on all five attempts.
The result supports further investigation into how the software around a model affects its performance.
¹ Published comparison captured in the report. Backboard’s result is an internal evaluation and had not undergone leaderboard review at publication.
Exploring the cost of capability.
In a separate experiment, R-CLI used Kimi K3 to direct the work and DeepSeek V4 Flash for delegated tasks.
91.0%
task completion. 81 of 89 tasks solved, with one attempt per task.
$28.72
in recorded model charges. Provider-reported charges across the available usage records.
This experiment explores how model selection, context reuse, and execution controls can work together within a limited budget.
Exploratory result, not a reviewed leaderboard score. One attempt per task is not directly equivalent to the five-attempt Opus evaluation. Charges exclude sandbox infrastructure and omit one trial with missing usage telemetry.
Inside the agent.
Coordinate the work.
R-CLI can use different models for primary and delegated work. This lets us investigate which tasks need a more capable model and where a less expensive model can contribute.
Manage the context.
Core instructions and tool definitions remain stable, while additional capabilities enter context when needed. We study how selecting and reusing information affects task completion and resource consumption.
Control execution.
Safe, independent tool calls can run together. Actions requiring ordering run sequentially. Time budgets and spending signals help the agent prioritize its remaining work.
Make failures inspectable.
Execution traces record model interactions, tool calls, errors, and available usage data. These records help us examine where the agent succeeds, where it stops too early, and where additional work fails to improve the outcome.
Results with their limits.
These studies measure complete agent configurations. They do not isolate the contribution of each architectural choice or establish performance across every coding workflow.
Both reports include configurations, per-task results, failure analysis, and evaluation limitations. Neither result had completed leaderboard review at publication. The Opus report also documents a failed sandbox restart barrier in 444 of 445 trials.
The next research steps are repeated evaluation, external review, and experiments that measure individual components.
From the blog
- Aug 2026BlogWe Scored 91% on Terminal-Bench 2.1 With Kimi. That’s Not the Interesting Part.Read
- Aug 2026BlogWe Open Sourced the Harness That Beat Claude Code and CodexRead
- Aug 2026BlogWe Beat Claude Code With ClaudeRead
- Jul 2026BlogBackboard R-CLI 3.0.2: Faster Workflows, More ControlRead
- Jul 2026BlogBackboard R-CLI Is Now the Top Result on Terminal-Bench 2.1Read