{"slug": "show-hn-pantheon-ai-vs-ai-one-writes-the-code-the-other-attacks-it", "title": "Show HN: Pantheon – AI vs AI: one writes the code, the other attacks it", "summary": "A developer released Pantheon, an open-source multi-agent harness for Claude Code that runs coding tasks through parallel implementations and adversarial verification to catch bugs that single-pass models miss. The tool uses a pipeline of planning, parallel implementation, adversarial testing by a second AI, and synthesis, and is available in two versions: pantheon (Claude-only) and pantheon-x (cross-model with GPT-5.5).", "body_md": "Two Claude Code skills that run a hard coding task through a multi-agent harness instead of a single model pass: **plan → N parallel implementations → adversarial verification → judge**. The point isn't a smarter model — it's that a second (and third) implementation, plus an independent reviewer whose job is to *break* the result, catches bugs a single pass ships green.\n\nIt's a packaging of well-worn techniques — best-of-N sampling, tool-integrated self-correction, and LLM-as-judge / adversarial verification — wired into one `/pantheon`\n\ncommand so you don't reassemble them by hand each time. This is scaffolding *around* the model, not a change *to* it: it won't rescue a task the model fundamentally can't reason about, but it reliably tightens correctness on coding work whose answer you can express as tests.\n\nThe harness runs a deterministic pipeline:\n\n```\nPlan ──▶ Implement (×N parallel) ──▶ Verify (adversarial ×V) ──▶ Synthesize\n │            │ each self-corrects            │ try to BREAK each      │ judge picks winner\n 1 planner    │ against its own tests (T1)    │ green build            │ + grafts best ideas\n              N builders                       reviewers\n```\n\n**Plan**— derive a tight spec, a test plan that*defines*correctness, and N distinct strategies (before any code).**Implement**— N builders implement different strategies in parallel; each runs its own tests and self-corrects on failure (tool-integrated self-verification, up to 5 iterations).**Verify**— independent adversarial reviewers try to*break*each green build; a build refuted by a majority is dropped.**Synthesize**— a judge picks the winner and lists superior ideas worth grafting from the runners-up.\n\nThe value: a build can pass its *own* tests yet still be wrong. The adversarial layer catches defects the self-written tests miss, instead of rubber-stamping a green build.\n\n| Skill | Adversarial verifier | Requirements |\n|---|---|---|\n`pantheon` |\nClaude itself (independent agents) | Paid Claude Code plan + Workflows (see below) |\n`pantheon-x` |\nGPT-5.5 via Codex plugin (cross-model) |\nAbove + OpenAI Codex plugin (`codex:codex-rescue` ) |\n\n`pantheon-x`\n\nis the stronger setting: the implementation written by Claude is attacked by a *different* model, which shrinks single-model blind spots (the same mistake slipping past a same-model verifier). If you don't have Codex/GPT-5.5, use `pantheon`\n\n.\n\nBoth skills share the same harness (`pantheon-class.js`\n\n); they differ only in the `crossModelVerify`\n\nflag.\n\nThese skills drive Claude Code's **Workflow** orchestration engine, so a stock/Free setup is not enough:\n\n**Claude Code ≥ v2.1.154** on a**paid plan**— Pro, Max, Team, or Enterprise (also Bedrock / Vertex / Foundry).** Not available on the Free tier.**- On\n**Pro**, enable it once:`/config`\n\n→ turn on**Dynamic workflows**. the cross-model verifier runs as the`pantheon-x`\n\nonly:`codex:codex-rescue`\n\nsubagent, which ships in OpenAI's**Codex plugin**—*not*stock Claude Code. A logged-in`codex`\n\nCLI alone does**not** register it. Install the plugin:plus a ChatGPT subscription (or\n\n```\n/plugin marketplace add openai/codex-plugin-cc\n/plugin install codex@openai-codex\n```\n\n`OPENAI_API_KEY`\n\n) and the`codex`\n\nCLI on PATH.**If**—`codex:codex-rescue`\n\nisn't installed, use`pantheon`\n\ninstead`pantheon-x`\n\nwould otherwise silently skip the adversarial pass and pass every build.\n\nSkills and subagents themselves are stock Claude Code features; no extra setup beyond the above.\n\nClone into your Claude Code skills directory (personal install):\n\n```\ngit clone https://github.com/lolu1032/pantheon-skills.git\ncp -R pantheon-skills/pantheon       ~/.claude/skills/pantheon\ncp -R pantheon-skills/pantheon-x     ~/.claude/skills/pantheon-x\n```\n\nOr for a single project, copy into `<project>/.claude/skills/`\n\n.\n\nIn Claude Code:\n\n```\n/pantheon    <a hard implementation task whose correctness is testable>\n/pantheon-x  <same, but GPT-5.5 does the adversarial verification>\n```\n\nExample:\n\n```\n/pantheon Add idempotency-key handling to the payments module so concurrent requests can't double-charge. Tests: pnpm test (vitest)\n```\n\nClaude collects the parameters (`task`\n\n, `workdir`\n\n, `lang`\n\n+ test command, `variants`\n\n, `verifiers`\n\n) and launches the harness as a background Workflow, then reports: per-variant test results, which builds the adversarial pass broke, and the final winner with its rationale and grafting suggestions.\n\n| arg | default | notes |\n|---|---|---|\n`task` |\n— | one-paragraph requirement + acceptance criteria (expressible as tests) |\n`workdir` |\n`/tmp/pantheon-<name>` |\nabsolute path; a real repo or a scratch dir |\n`lang` |\nPython/unittest | language + the exact test command for your stack |\n`variants` |\n3 | bump to 5 for harder problems |\n`verifiers` |\n2 | bump to 3 to be stricter (majority refutation drops a build) |\n`crossModelVerify` |\n`false` (`pantheon` ) / `true` (`pantheon-x` ) |\nroute adversarial verify to GPT-5.5/Codex |\n\n**Not a daemon.** Each invocation runs once to completion and exits — zero cost when idle.- A run spends real tokens. A representative run is ~11 subagents and a few hundred K to ~1M tokens end-to-end, ~6–10 min wall-clock; heavier settings (\n`variants=5`\n\n,`verifiers=3`\n\n, cross-model) cost more. On Pro/Max it draws from your usage quota; on metered API access, budget a few dollars per run and up.**Route only the hardest 10–20% of tasks here**— use plain Opus for the rest. - This buys\n*correctness on testable work*, not raw model intelligence. If a task isn't expressible as tests, the adversarial layer has little to grip and the overhead isn't worth it. - Coding/agentic productivity only.\n**Not** a tool for bypassing safety gates (cybersecurity/biology capability restrictions).\n\n**Isn't this just a prompt wrapper?**\nThere's no model change — it's orchestration, yes. The non-trivial part is the *adversarial* step: an independent agent (a different model in `pantheon-x`\n\n) whose job is to break a build rather than confirm it. That's what catches defects the builder's own green tests rubber-stamp. The value is the harness shape, not a secret prompt.\n\n**Do you have benchmarks vs. plain Opus?**\nNo formal benchmark yet — treat the description as *mechanism*, not a measured delta. The value is in the adversarial step: a build can pass its own tests and still be wrong, and an independent reviewer catches what the self-written tests rubber-stamp. If you run a head-to-head, I'd genuinely like to see the numbers.\n\n**What does a run cost?**\nA few hundred K to ~1M tokens and ~6–10 min at default settings; more for `variants=5`\n\n/ `verifiers=3`\n\n/ cross-model. It's meant for the hardest 10–20% of tasks, not everyday edits. See [Cost & scope](#cost--scope).\n\n**It says \"Workflow tool not found\" / nothing happens.**\nYou're likely on the Free tier, or haven't enabled workflows. See [Requirements](#requirements) — needs a paid plan and, on Pro, `/config`\n\n→ **Dynamic workflows**.\n\n**Why route verification to GPT-5.5 / another vendor's model?**\nSame-model verifiers share blind spots — a mistake the builder makes, a same-model reviewer tends to miss too. A *different* model is a cheap way to break that correlation. It's optional: `pantheon`\n\nruns Claude-on-Claude and still helps.\n\nSolo project, **as-is, best-effort**. Issues and PRs are welcome, but maintenance comes with no guarantees or SLA — I may not get to everything. It's MIT-licensed, so forking is a first-class option if you want to take it further.", "url": "https://wpnews.pro/news/show-hn-pantheon-ai-vs-ai-one-writes-the-code-the-other-attacks-it", "canonical_source": "https://github.com/lolu1032/pantheon-skills", "published_at": "2026-06-15 14:20:33+00:00", "updated_at": "2026-06-15 14:38:42.706950+00:00", "lang": "en", "topics": ["ai-agents", "developer-tools", "large-language-models", "ai-safety"], "entities": ["Claude Code", "OpenAI", "GPT-5.5", "Codex", "Pantheon"], "alternates": {"html": "https://wpnews.pro/news/show-hn-pantheon-ai-vs-ai-one-writes-the-code-the-other-attacks-it", "markdown": "https://wpnews.pro/news/show-hn-pantheon-ai-vs-ai-one-writes-the-code-the-other-attacks-it.md", "text": "https://wpnews.pro/news/show-hn-pantheon-ai-vs-ai-one-writes-the-code-the-other-attacks-it.txt", "jsonld": "https://wpnews.pro/news/show-hn-pantheon-ai-vs-ai-one-writes-the-code-the-other-attacks-it.jsonld"}}