Agent code execution

에이전트 코드 실행

Run agent-generated code in an isolated environment and feed the result back into the task.

···
html
<div class="scene"><div class="heading">AGENT → SANDBOX → RESULT</div><div class="sandbox"><div class="code-line mono">sum([4, 7, 9])</div><div class="run-line" id="run-line">대기 중</div><div class="output-line mono" id="code-result">result: —</div></div><div class="muted" id="code-note">코드를 격리된 환경에 전달</div></div>
css
.scene{width:min(94vw,800px);height:min(88vh,326px);padding:clamp(10px,2.7vmin,19px);border:1px solid var(--line);border-radius:13px;background:var(--surface);font:500 clamp(13px,2.4vmin,17px)/1.3 var(--font-sans,sans-serif);position:relative;overflow:hidden}.scene .mono{font-family:ui-monospace,SFMono-Regular,monospace}.scene .muted{color:var(--muted)}.scene .accent{color:var(--accent)}.scene .heading{font-weight:750;color:var(--accent);margin-bottom:clamp(5px,1.6vmin,12px)}.sandbox{border:2px dashed var(--accent);border-radius:8px;padding:clamp(9px,2.5vmin,18px);height:58%;display:flex;flex-direction:column;justify-content:space-between;background:var(--bg);margin-bottom:6px}.code-line{color:var(--accent)}.run-line{height:clamp(17px,4vmin,25px);border:1px solid var(--line);border-radius:4px;padding:0 7px}.run-line.active{background:color-mix(in srgb,var(--accent) 18%,var(--surface));border-color:var(--accent)}.output-line{color:var(--muted)}.output-line.done{color:var(--accent);font-weight:700}
js
const states=[['대기 중','result: —','코드를 격리된 환경에 전달'],['실행 중','result: —','샌드박스에서 계산'],['완료','result: 20','실행 결과를 에이전트가 확인']];let step=0;function draw(){const [status,result,note]=states[step];document.getElementById('run-line').textContent=status;document.getElementById('run-line').classList.toggle('active',step===1);document.getElementById('code-result').textContent=result;document.getElementById('code-result').classList.toggle('done',step===2);document.getElementById('code-note').textContent=note;step=(step+1)%3}draw();setInterval(draw,900)

A code execution tool runs a short model-written program in a sandbox and returns stdout or generated files. It supports calculation, data transforms, and visualization where execution is more reliable than guessing text.

The demo runs a sum inside a visible boundary and returns the result to the agent. Set limits on execution, files, and time, and validate the returned result.

When to use

Use it when calculations or data processing should be executed instead of inferred in prose.

Open as page ↗