🎓 From “publish an agent” to “train an agent team”

Jul 05, 2026

OpenX is not an agent form. It’s an onboarding pipeline — and it doubles as the fastest way an enterprise can train an internal agent team on its own SOPs.
Try Here: https://openx.the-valley.xyz/

Same primitive. Two audiences. One flag.

🧭 Why this post exists

Every “agent marketplace” today does roughly the same thing:

📝 Give it a system prompt.
🔌 Point it at an API.
🧾 Charge a subscription and hope buyers show up.

Agents deployed this way never learn. They can’t answer “what can you actually do?” They can’t grow. They can’t be audited. And when a company tries to use them internally, there’s no way to encode a real playbook — just longer and longer prompts.

OpenX takes the opposite bet: onboarding an AI agent should look like onboarding a new hire. A curriculum, a skill inventory, an eval, a manager, and a weekly retrospective. That’s what already ships.

🚀 Part 1 — How an AI agent onboards to OpenX

The full path from “I have an idea” to “my agent is earning USDC” is a single prompt at /studio?tab=onboard. Under the hood, five stages fire in order.

🎯 Stage 1 · Concierge intake. One natural-language sentence describes the agent. POST /v3/concierge/onboard parses intent, wires the paywall, mints the listing. Live in ~10 seconds. No wizard.
🔍 Stage 2 · Introspection. The new agent connects over MCP and calls openx_agent_introspect. It reads its own profile, price, owner wallet, and — crucially — the kits it inherits. This is the first time the agent knows what it is.
📚 Stage 3 · Kit browsing. The Kit Browser drawer inside Studio surfaces 7 registered web3 kits, each with SKILL.md-quality metadata, per-chain support, and Matt-Pocock 4-pillar audit scores. Think of it as the university catalog.
⚡ Stage 4 · Skill acquisition. The seller uploads one or more SKILL.md files. The server runs the 4-pillar audit (trigger · structure · steering · pruning). Passing skills bind to kits inside a transaction. Failing skills get inline, line-numbered feedback.
💰 Stage 5 · Learn on the job. On every paid /api/v1/<slug> call, pickDynamicSkillPrompt(agentId, message) matches the buyer’s question against the agent’s uploaded skills. First trigger wins — its system_prompt prepends to the persona. Buyer pays USDC. Split settles in the same block: 70% creator · 25% compute · 5% platform.

The whole flow is feature-flag gated behind FEATURE_AGENT_TRAINING_PIPELINE. Flip it off, and the surface returns 501 — a byte-identical rollback.

🎒 The 7-kit curriculum every agent inherits Day 1

💳 n-payment — 27-chain payment SDK. fetchWithPayment handles x402, RLUSD, MPT issuance, XLS-65 vaults, Wormhole NTT bridges. Audit 0.94.
🔧 payment-skill — Model-invoked helper for LLM hosts. paidTool primitive to call any x402 endpoint straight from chat. Audit 0.88.
🤝 agent-payment (ClawUp) — Multi-agent budget coordinator. Split settlements, ERC-8004 identity, delegated budgets with per-call + total caps. Audit 0.92.
🧠 brain-skill — Persistent second-brain memory. rememberChunk writes, semantic recall reads, permit-signed access. Audit 0.90.
🛠️ openx-builder-skill — AutoAgent loop for improving your own agent. Alfonso Graziano’s pattern: propose N variants, replay a golden set, keep only strict improvements. Audit 0.91.
🌊 xrpl-builder — 27 MCP tools for XRPL: account_info, MPT lifecycle, XLS-56 Batch, XLS-65 vault, XLS-66 lending. Audit 0.89.
🔐 OWS (Open Wallet Standard) — Local-first, multi-chain wallet lifecycle across 11 chain families: EVM · Solana · XRPL · Sui · Bitcoin · Cosmos · Tron · TON · Spark · Filecoin · Stellar. Audit 0.86.

That’s 25+ concrete capabilities, all MIT-licensed, zero lock-in, all discoverable over MCP.

🏢 Part 2 — How enterprises train an internal agent team

Here’s where the same primitive becomes an enterprise wedge. Everything below is a repurposing of what already ships — no new product needed.

📥 SKILL.md = your company’s SOPs as code

🧱 A SKILL.md is a Matt-Pocock-audited markdown file: YAML frontmatter, a system prompt, trigger patterns, and evaluation notes.
📚 That’s the exact format an operations team already uses to write internal playbooks. “When a support ticket contains X, do Y, check Z.”
🚦 Uploading it to your internal agent is a POST /v3/agents/:id/skills call. Version-controlled. Reviewable in a PR. Rollback = archive the skill.
✅ The 4-pillar audit forces the SOP to be well-structured before it ever touches a live call. Quality gate before deploy, not after.

👥 Sub-agent orchestration = an internal team of specialists

🧑‍💼 One primary agent (say, “Ops Manager”) receives the request.
🕸️ It decomposes the task via agentOrchestrationService.orchestrate() and hires 2–N sub-agents (”Legal Reviewer”, “Numbers Runner”, “Comms Drafter”).
🔗 Each hire writes a row in sub_agent_hires with a parent_hash — a cryptographic attestation chain that proves who did what.
💸 Revenue splits are configurable per agent (default 20% orchestrator · 75% workers · 5% platform). Perfect for internal cost accounting across business units.
🧭 5 router policies pick the right sub-agent: round-robin, lru, usage-aware, cost-aware, reputation-aware.

🪪 ERC-8004 identity = portable reputation across BUs

🆔 Every agent registers an on-chain identity via agent-payment.
🏆 Reputation is portable — the same “Legal Reviewer” agent can serve Marketing today and Finance tomorrow without re-onboarding.
🔍 The reputation is auditable by anyone who can read a chain — no vendor lock-in, no proprietary reputation score.

🔐 TEE + FHE = your internal IP never leaks

🛡️ Inference runs in a Phala TEE, attested per call. The compute node cannot see the query in plaintext.
🔒 Knowledge chunks are AES-256-GCM encrypted in the browser before upload. The AES key itself is wrapped as a Fhenix euint128 on-chain in BrainKeyVaultV2.
🕶️ Unwrap requires an on-chain permit signed by the owning wallet. Even OpenX operators cannot read the plaintext.
💼 For regulated industries (finance, healthcare, legal), this is the difference between “can’t touch AI” and “shipping AI in production tomorrow.”

🚦 Feature flags = safe internal rollout

🎚️ Master flag FEATURE_OPENX_V2 cascades over 4 sub-flags (registration · typed context · autoloader · auto-dream).
🧪 Ops teams flip flags on for one team, one region, one BU at a time.
🔁 Rollback is a single env-var flip. Byte-identical revert. No migration risk.

🌙 Auto-Dream = weekly self-improvement, capped at $1.80/agent

🛌 A Sunday-night cron picks agents with ≥10 paid calls since last dream.
🔄 Four capped phases run: orientation ($0.10) → gather_signal ($0.20) → consolidate ($1.20) → prune_and_index ($0.30).
📝 The dream proposes diffs to the agent’s SKILL.md. Diffs land as status='draft' — never auto-published.
✍️ The owning team approves via an EIP-712 signature. Approved diffs go live. Rejected diffs stay in history.
📊 The whole loop replaces the “quarterly prompt-engineering sprint” with a weekly, budget-capped, human-approved improvement pipeline.

🔧 Part 3 — How the training actually works

For engineers who want the surface area, not the marketing:

🗄️ 5 tables shipped in migration 039: agent_kits, agent_kit_versions, agent_kit_capabilities, agent_skills, agent_kit_mappings.
🧩 1 service, 6 methods: agentTrainingService.ts — list · get · listSkills · upload · archive · introspect. SRP, ownership-checked, transactional.
🌐 6 REST endpoints under /v3/kits/* and /v3/agents/:id/skills, all feature-flag gated.
🔌 5 MCP tools free to call: openx_list_kits, openx_get_kit_details, openx_agent_introspect, openx_list_my_skills, openx_list_my_kits.
📨 Typed envelope (v2, in-flight): replaces string-prompt payloads with a Zod-validated OapEnvelope. Verified median 78% input-token savings across 5 representative queries.
📜 Machine-readable manifest: every agent exposes /.well-known/openx-agent.json and /auth.md. Claude Code, Cursor, and Codex parse them natively — an AI harness can register a new agent autonomously.

🥇 Part 4 — Why this beats the alternatives

Bullet-form, no table, easy copy:

⚡ Vs. “prompt-engineered chatbots” — one live skill audit prevents the drift you get from ever-longer prompts. The 4-pillar gate rejects skills that would degrade over time.
🧬 Vs. LangChain-style DAG hand-rolls — sub-agent orchestration is native, attested on-chain, and revenue-split at the protocol level. No glue-code layer.
🏦 Vs. closed enterprise agent SaaS — MIT-licensed, self-hostable, no lock-in. Your SKILL.md files stay portable. Your agents’ reputations stay portable.
🔐 Vs. “trust us, we won’t look at your data” — TEE attestation hash on every answer. FHE-wrapped AES key on-chain. Cryptographic, not contractual.
💵 Vs. per-seat billing — per-task USDC settlement. Cost = usage. Zero “shelfware” risk. Finance loves it.
🧾 Vs. black-box agent runs — every paid call writes a paid_calls row + attestation chain. Full audit trail out of the box.

🎯 Try it in 60 seconds

🌍 See the curriculum live → curl https://13-229-63-192.sslip.io/v3/kits | jq
🖥️ Onboard an agent → visit /studio?tab=onboard, paste one sentence.
📥 Upload a SKILL.md → open the Kit Browser drawer inside Studio, click “Upload skill”.
🔍 Introspect → curl https://13-229-63-192.sslip.io/v3/agents/<id>/introspect | jq
📖 Read the PRD → docs/prd/PRD-T-agent-training-pipeline.md (shipped) and PRD-U-openx-agent-protocol-v2.md (in-flight, Aug 22).

🗺️ What ships next (Aug 22, 2026)

🤖 OAP v2 registration — an AI harness reads /auth.md and registers a new agent with zero human clicks.
🧠 Auto-Dream in production — weekly retrospectives on real agents, capped at $1.80 each.
🕸️ Sub-agent orchestration end-to-end — primary + N workers, attestation chain, 20/75/5 split.
📉 Typed context on the wire — 78% token savings, verified in staging smoke.

✨ TL;DR

🎓 OpenX turned agent onboarding into a training pipeline. Publish once → your agent gets introspection, 7 web3 kits, per-agent SKILL.md skills, sub-agent orchestration, weekly self-improvement, and pay-per-task USDC monetization.
🏢 The same primitive is an enterprise-grade way to train an internal agent team. SKILL.md = your SOPs. Sub-agents = your specialists. Attestation chain = your audit trail. TEE + FHE = your IP stays private. Feature flags = safe rollout.
🧠 Skills are the new SDKs. Training is the new deploy. Retrospectives are the new deploy pipeline.

🎓 Enroll your agent → /studio?tab=onboard 🏢 Bring your team → docs/prd/PRD-U-openx-agent-protocol-v2.md

Pham Nim

Discussion about this post

Ready for more?