The Coding Agent Leaderboard Is a Photo Finish. Pick the Harness.

The useful conclusion from the current coding-agent race is not that every tool is interchangeable. It is that a headline score now says less about the thing you will use at 9:12 a.m.: the model, agent harness, benchmark mix, operating system, permission model, and route into code review are all part of the result. Here are the developments that make tool selection more operational than leaderboard-driven.

Composite leaderboards now admit that one score hides different work

Artificial Analysis’s current Coding Agent Index combines DeepSWE, Terminal-Bench v2, and SWE-Atlas-QnA, rather than treating one bug-fix benchmark as a complete verdict. Its own methodology explicitly says to inspect per-evaluation results: agents with similar composite scores can diverge between repository work, terminal workflows, and rubric-scored technical questions. [1]

Why it matters: stop asking which agent is “best” and run a small acceptance suite in the tool you would actually deploy—one repo issue, one shell-heavy maintenance task, and one code-reading task—with your normal test command and review rules. A composite is useful for eliminating clearly weaker options; it is bad at choosing between close contenders for a particular codebase.

Fresh, multilingual, and Windows tasks are exposing old benchmark blind spots

SWE-bench-Live is now an automatically updated task set, and its February 2026 update added C, C++, C#, Java, Go, JavaScript/TypeScript, Rust, and Windows coverage. The project notes that SWE-agent, OpenHands, and Claude Code could not run in its Windows containers, so it created a minimal Windows-compatible harness for those evaluations; its original Lite, Full, and Verified splits remain Python-only. [2]

Why it matters: a strong Python-on-Linux score is not evidence that an agent can operate in your PowerShell setup, build your Java monorepo, or even start in the required environment. If your team is outside the familiar Python/Linux path, environment support should be a purchase gate before model quality becomes a tiebreaker.

The harness has become a measurable part of agent capability

The same leaderboard now holds the underlying model constant—Claude Opus 4.7 in its published harness comparison—and compares Claude Code, Cursor CLI, and OpenCode. That is an important shift: context construction, tool definitions, retry behavior, permission interruptions, and the way a harness manages a long task can change the outcome even when the model is identical. [1]

Why it matters: choosing an agent is increasingly choosing its control loop, not merely choosing access to a model. Try the concrete failure cases: an interrupted test run, a task that needs a second working directory, a command that needs network access, and a request to produce a minimal diff; those are precisely where a polished benchmark patch can turn into a frustrating daily interaction.

The product race is moving from one agent to coordinating several

OpenAI’s Codex app added a desktop command center organized around parallel agents and isolated worktrees, with reviewable diffs and project rules for elevated commands; it became available on Windows on March 4, 2026. Google made the same strategic move in May, transitioning Gemini CLI users to Antigravity CLI and saying the CLI and Antigravity 2.0 desktop app would share an agent harness; consumer Gemini CLI service ended June 18, 2026. [3] [4]

Why it matters: if you regularly delegate more than one issue, compare worktree isolation, task visibility, handoff behavior, and the review queue—not just single-agent patch quality. The downside is real: a switch from a familiar CLI to a new coordinated-agent surface is workflow churn, so verify export, local-repository behavior, and team rollout before rebuilding scripts around it.

Repository-owned agent configuration is becoming the safer differentiator

GitHub now lets teams define custom agent profiles as Markdown files, with a prompt, selected tools, model choice, and MCP-server configuration at repository, organization, or enterprise scope. A profile can, for example, expose only read, edit, search, and shell execution to a bug-fixer; GitHub also warns that some configuration properties differ or are ignored between cloud, CLI, and IDE environments. [5]

Why it matters: when the leaders are close, choose the tool that lets you commit the team’s operating procedure alongside the code—how to run tests, which directories are off-limits, which MCP tools are available, and what a finished task must report. This is less glamorous than chasing a new top score, but it is how an agent becomes repeatable across developers and reviewable by the people who inherit its changes.

Permission controls remain more valuable than another point on a chart

Claude Code’s CLI still makes the practical controls explicit: claude -p --max-turns 3 "…" can bound a non-interactive run, --permission-mode plan starts it without broad execution, and --dangerously-skip-permissions exists precisely because bypassing approval is risky. Codex similarly defaults to limiting agents to the working folder or branch and cached web search, requesting permission for elevated actions such as network access. [3] [6]

Why it matters: an agent that needs a little more steering but leaves a comprehensible diff is often the better everyday tool than the one that completes a benchmark task with fewer turns. Pick a default permission posture first, then compare capability within that boundary; otherwise the leaderboard is measuring an autonomy level your team may never be willing to enable.