# Corgent *This page is a high-level overview. Details, APIs, and flows will move into sub-pages as Corgent evolves.* *** ### TL;DR **Corgent** is the **core Agentic AI Infra surface of Cortensor** – an **ERC-8004–native trust and execution agent** designed for **agent ecosystems in general** (8004 agents, agent frameworks, and protocol-level integrations). It exists so agents and dApps don’t have to blindly trust a single model run. Instead, they can call Corgent to: * **Delegate** tasks to Cortensor miners with explicit reliability tiers * **Validate** claimed results using **PoI / PoUW**, redundancy, and node reputation * **Arbitrate** conflicts using oracle-grade replays and consensus At a high level: > **Agent / dApp decides → Corgent delegates or validates → Cortensor computes & proves → Corgent returns a verdict → Agent / dApp continues.** Corgent is the **canonical, infra-native agent surface** of Cortensor.\ **Bardiel** is a **Virtual-first, product-layer agent** that *also supports ERC-8004*, but focuses on value-added features and UX built on top of this core Corgent surface. *** ### What Corgent Is Corgent is a **router-side oracle & policy layer** that exposes Cortensor’s decentralized AI stack to **ERC-8004 agents and agent ecosystems** in a clean, programmable way. Conceptually, Corgent: * Sits directly on top of **Router Nodes, Session, and Validation modules** * Uses **miners** for inference and **PoI / PoUW + reputation + integrity checks** for trust * Returns **structured, machine-readable outcomes**, such as: * `VALID` – consistent, useful, and within spec * `INVALID` – conflicts with consensus, schema, or constraints * `RETRY` – re-run with refined spec or stronger tier * `NEEDS_SPEC` – task is underspecified or structurally unclear From the ERC-8004 perspective, **Corgent is the base trust oracle & execution service** that any compliant agent, marketplace, or agent framework can call for **delegated inference, result verification, or dispute resolution**. *** ### What Corgent Is Not Corgent is **not**: * A **planner / goal generator** – higher-level agents or frameworks decide *what to do* * A **miner** – Cortensor miners perform the actual inference work * The **base protocol validator set** – on-chain validators enforce network rules and economics * A UX-heavy or “creative” product agent – Corgent focuses on **trust, policy, and orchestration**, not front-end ergonomics Think of Corgent as **core trust middleware** between agent ecosystems and the Cortensor network – minimal, stable, and designed to be built on. *** ### Where Corgent Lives in the Stack **Layering (for ERC-8004 & agent ecosystems):** 1. **Agents & Platforms** * ERC-8004 agents, registries, and marketplaces * Agent frameworks and orchestrators * Web2 backends, bots, and services that speak HTTP/x402 2. **Corgent (Core Trust & Orchestration Surface)** * Policy selection (Fast / Safe / Oracle-grade / Adaptive) * Delegation of tasks into Cortensor * Validation & arbitration logic * Verdicts + evidence back to callers (8004-compatible metadata) 3. **Cortensor Network** * Router Nodes & Session Queue * Miner Nodes (decentralized inference) * Validation: **Proof of Inference (PoI)**, **Proof of Useful Work (PoUW)**, node reputation, integrity checks * Payments, staking, and settlement on L2/L3 4. **Base Chains** * Testnet-0 (Arbitrum Sepolia L2) * Testnet-1 (COR L3 Rollup) * Future Mainnet-Lite (Arbitrum L2) * Future Mainnet-Full (COR L3) Corgent packages this stack into a **canonical, ERC-8004-ready interface** that other agents can trust and compose with. *** ### Core Capabilities #### 1. Delegation-as-a-Service > “Run this task on Cortensor with reliability tier X.” * Agents send **task + context** (prompt, inputs, constraints, spec) * Corgent selects **model, redundancy, and validation tier** * Submits the job into Cortensor via Router / Session * Aggregates miner outputs + PoI / PoUW / reputation signals * Returns an **oracle-validated output** plus optional confidence / evidence Use when you want **reliable execution** rather than a one-off opaque model call. *** #### 2. Validation-as-a-Service > “Verify that this other agent’s result is correct/useful.” * Agents send `{task, claimed_result, spec}` * Corgent chooses a validation pattern: * Deterministic → PoI rerun + direct comparison * Non-deterministic → N-of-M consensus * Structured → Schema/spec checks as a baseline * Triggers reruns / consensus on Cortensor as needed * Combines **PoI, PoUW, reputation, and integrity checks** * Returns `VALID / INVALID / RETRY / NEEDS_SPEC` + machine-friendly reasoning fields Use as a **trust layer, guardrail, and audit channel** around important actions. *** #### 3. Arbitration-as-a-Service > “Two parties disagree — give me a binding oracle view.” * Typical for buyer/seller disputes in **8004 marketplaces** or cross-agent conflicts * Corgent escalates to **oracle-grade policy** (more miners, diversity, strict thresholds) * Recomputes the task under tight, pre-declared rules * Compares disputed outputs vs **oracle consensus + spec** * Returns a verdict that downstream systems can treat as **binding** for settlement or governance Use when you need **replayable, evidence-backed dispute resolution**, not ad-hoc heuristics. *** ### Validation Tiers & Policy Corgent exposes **policy tiers** to balance cost vs assurance: * **Fast** * 1 miner * Light spec/schema + basic checks * Low latency, low cost * **Safe** * 3 miners * PoI consensus + basic usefulness scoring * Recommended default for most agent calls * **Oracle-Grade** * 5+ miners with diversity sampling * Strict PoI, richer PoUW rubrics, stake/reputation weighting * For high-stakes decisions and disputes * **Adaptive** * Start with low-cost run * Auto-escalate redundancy and validation depth when confidence is low These policies are intended to be **stable, documented, and versioned**, forming the backbone of **Agentic AI Infra** for ERC-8004 agents and other ecosystems. *** ### ERC-8004 Integration Corgent is designed to be a **first-class ERC-8004 citizen**: * Registered as an **8004-compatible service/agent**, discoverable by 8004 registries * Accepts **standardized agent task payloads** (with spec/metadata) * Returns **machine-readable verdicts and artifacts** that can be referenced by: * Other agents in their decision logic * Marketplaces and settlement layers * On-chain verification or scoring systems Typical patterns: * Offload heavy or sensitive tasks from an 8004 agent to Corgent for **delegated execution** * Validate another agent’s output before **triggering payments or state changes** * Use arbitration flows for disputes in **8004-based marketplaces** or long-running agent workflows Sub-pages will define the **exact 8004-aligned payloads, schemas, and examples**. *** ### Corgent vs Bardiel (Positioning) Both **Corgent and Bardiel support ERC-8004**, but **they serve different roles**: * **Corgent – Core Agentic AI Infra (Infra-Native)** * Primary, **canonical agent surface** of the Cortensor Network * Focused on **core agent services**: delegation, validation, arbitration * Tight alignment with Router, Sessions, and validation modules * Interface designed to be **lean, stable, and composable** for protocols and frameworks * The “base layer” other agents (including Bardiel) can standardize on * **Bardiel – Value-Added Ecosystem Agent (Virtual-First + ERC-8004)** * **Virtual-native service agent**, built on Cortensor and often on Corgent * Also capable of registering and operating in **ERC-8004 ecosystems** * Focused on **product-level and UX features**: * GAME/ACP-friendly flows * Ecosystem-specific presets and policies * Potential token mechanics, higher-level workflows, and trajectory tooling * Evolves faster at the edges to match **agent demand and UX needs** You can think of it as: > **Corgent** = *Core Agentic AI Infra* – the canonical “agent trust surface” for Cortensor.\ > **Bardiel** = *Value-added product agent* – built on that surface, specialized for Virtual and richer UX, but also able to participate in the ERC-8004 world. In practice: * If you are designing **protocols, infra, or 8004-native agents**, you generally integrate **Corgent** as your primary trust/execution layer. * If you are building **Virtual-first experiences** or need richer UX and ecosystem integrations, you may integrate **Bardiel**, which in turn leans on Corgent/Cortensor underneath. *** ### Relationship to Router Node Evolution Corgent builds directly on the **Router Node roadmap**: * **Router v1.5** – Adds MCP/x402-style surfaces and stable `/completions` endpoints * **Router v1.6** – Adds `/validate` and completes the **Delegate + Validate** pair * **Corgent v1.x** – Wraps these into **policy-driven, ERC-8004-ready Agentic AI Infra services** Over time, Corgent becomes the **canonical “agent interface” for Router + Validation**, while Bardiel remains a **higher-level ecosystem agent** that can adopt new product features faster. *** ### Why Corgent Matters For **agent ecosystems & ERC-8004**: * Provides a **neutral, reusable trust oracle** that any agent can call * Makes **verifiable inference and validation** a core primitive instead of an add-on * Helps agents avoid lock-in to a single model or vendor by using **multi-miner consensus + PoI/PoUW** For the **Cortensor Network**: * Turns raw decentralized compute + validation into a **clean, agent-ready contract** * Anchors Cortensor in **real agent workflows**, not just human-facing APIs * Defines a long-lived **Agentic AI Infra layer** that other products (like Bardiel) can build on Crucially, Corgent is **not meant to replace agent reasoning (GAME logic, planner LLMs, etc.)**. Instead: * **Agents still think; Corgent helps them be sure.** * Agents can call Corgent whenever they need **high-confidence decisions, verification for actions, or oracle-grade confirmation** using Cortensor’s redundant miners and PoI/PoUW.