Spaces:
Running
Running
| <html lang="en"> | |
| <head> | |
| <meta charset="utf-8"> | |
| <meta name="viewport" content="width=device-width, initial-scale=1"> | |
| <title>Why this stack — Agent Reliability</title> | |
| <style> | |
| :root { --bg: #1a1a1a; --fg: #e8e6e1; --muted: #9a9690; --accent: #d4a853; --card: #232323; --border: #353535; } | |
| * { box-sizing: border-box; } | |
| body { font: 17px/1.6 -apple-system, BlinkMacSystemFont, 'SF Pro Text', sans-serif; background: var(--bg); color: var(--fg); margin: 0; padding: 2rem 1.25rem; } | |
| main { max-width: 720px; margin: 0 auto; } | |
| h1 { color: var(--accent); font-size: 2.1rem; margin: 0 0 0.5rem; } | |
| h2 { color: var(--accent); font-size: 1.4rem; margin: 2.2rem 0 0.8rem; } | |
| .lede { color: var(--muted); font-size: 1.1rem; } | |
| blockquote { border-left: 3px solid var(--accent); padding: 0.4rem 1rem; margin: 1.2rem 0; color: var(--muted); font-style: italic; } | |
| ul li { margin-bottom: 0.4rem; } | |
| code { background: var(--card); padding: 0.1rem 0.4rem; border-radius: 4px; font-family: ui-monospace, SF Mono, monospace; font-size: 0.88em; color: var(--accent); } | |
| a { color: var(--accent); } | |
| footer { color: var(--muted); margin-top: 3rem; padding-top: 1.5rem; border-top: 1px solid var(--border); font-size: 0.9rem; } | |
| </style> | |
| </head> | |
| <body> | |
| <main> | |
| <h1>🧭 Why this stack</h1> | |
| <p class="lede">An honest take on what's actually broken about long-running AI agents — and why fixing it doesn't need a framework.</p> | |
| <h2>The problem isn't agents. It's the boring parts.</h2> | |
| <p>Most agent failures aren't reasoning failures. They're plumbing failures:</p> | |
| <ul> | |
| <li>Context window blew up halfway through a long task.</li> | |
| <li>The model called a tool with the wrong arg shape and the run died.</li> | |
| <li>The model "answered" but you can't parse the JSON.</li> | |
| <li>A tool quietly fetched something it shouldn't have, and you only noticed in the bill.</li> | |
| <li>You changed a prompt, output looks fine, but a downstream tool is silently using stale data.</li> | |
| </ul> | |
| <p>None of this is glamorous. None of it is novel research. All of it is wrecking your agents in production right now.</p> | |
| <h2>Five small fixes</h2> | |
| <blockquote>fit · guard · snap · vet · cast</blockquote> | |
| <ul> | |
| <li><strong>fit</strong> — fit messages into a token budget so the run doesn't 4xx</li> | |
| <li><strong>guard</strong> — declarative URL allowlist so tools can't fetch what they shouldn't</li> | |
| <li><strong>snap</strong> — snapshot tool-call traces so you catch silent regressions</li> | |
| <li><strong>vet</strong> — validate tool args before execution, return retry hints to the model</li> | |
| <li><strong>cast</strong> — extract + validate JSON from messy LLM output</li> | |
| </ul> | |
| <h2>What it isn't</h2> | |
| <p>This isn't a framework. There's no orchestration layer. No vendor lock-in. Each library is one job, ~200 lines, no runtime dependencies. Use one or use all five. Drop it into LangChain or LlamaIndex or your own loop — doesn't care.</p> | |
| <h2>BYO LLM</h2> | |
| <p>Nothing in here calls a model. <code>vet</code> doesn't know what an LLM is. <code>fit</code> doesn't either — it just counts tokens. <code>cast</code> assumes you've already gotten a string back from somewhere. You wire it up.</p> | |
| <h2>Why MIT, why no sponsorship</h2> | |
| <p>MIT, all the way down — Python, JavaScript, MCP servers, datasets. No commercial tier. No paid features. Take it, fork it, embed it in a closed-source product if you want.</p> | |
| <p>Sponsor links are intentionally absent — the author is on H1B, taking outside payments isn't allowed. Best contribution is filing an issue when something breaks, or starring the repos.</p> | |
| <h2>Try it</h2> | |
| <ul> | |
| <li><a href="https://huggingface.co/spaces/mukunda1729/agent-stack-demo">Live demo</a> — all 5 libraries in one Space</li> | |
| <li><a href="https://mukundakatta.github.io/agent-stack/">Landing page</a> — install commands + links</li> | |
| <li><a href="https://github.com/MukundaKatta">GitHub</a> — source for everything</li> | |
| </ul> | |
| <footer> | |
| Built by <a href="https://github.com/MukundaKatta">Mukunda Katta</a> · MIT licensed | |
| </footer> | |
| </main> | |
| </body> | |
| </html> | |