Corgent 2026 Roadmap¶
Last updated: 2026-06-28
Corgent is the agent-facing trust and execution layer built on Cortensor. The 2026 roadmap frames it as a Virtual-native trust oracle that can delegate work, validate outputs, escalate disputes, and produce evidence for agents and protocols.
What Corgent Adds¶
Delegation-as-a-Service
Agents send work to Cortensor miners with reliability policies, redundancy settings, and route/session selection.
Validation-as-a-Service
Agents check model outputs with PoI, PoUW, scoring, consensus, and evidence before acting.
Arbitration-as-a-Service
Disputes can escalate into stronger replay, validation, and oracle-grade review flows where enabled.
Virtual trust oracle
GAME, ACP, and Virtual agents get a common way to request high-trust compute and validation.
Corgent Timeline¶
Experiments and data foundations
GAME/ACP wiring, x402 prototypes, task/policy telemetry, early validation experiments, and trust recipes.
Corgent v0.1: spec and stable testnet agent
Delegate/validate API shape, policy tiers, basic schema validation, output format checks, and early testnet usage.
Corgent v0.2: validation engine
Tiered validation, PoI/PoUW scoring, outlier detection, confidence thresholds, dashboards, A/B testing, and evidence output.
Corgent v0.5: Virtual default trust oracle
GAME Worker/Function bundle, ACP soft disputes, high-trust recipes, usage limits, x402 billing paths, and production-oriented traffic.
Corgent v1.0: arbitration and long-horizon trust
Oracle-grade arbitration, multi-step trajectory validation, reusable trust artifacts, and COR L3 alignment.
Policy Tiers¶
| Tier | Meaning | Example fit |
|---|---|---|
| Fast | One miner with light checks and minimal overhead. | Low-risk content, inexpensive summaries, quick automation. |
| Safe | Three miners with PoI consensus and usefulness checks. | User-visible summaries, structured outputs, ordinary agent actions. |
| Oracle | Five or more miners, stricter thresholds, diversity sampling, and stronger validation. | Financial, on-chain, dispute, or high-trust workflows. |
| Adaptive | Starts cheaper and escalates when confidence is low. | Workflows where risk depends on task complexity or output quality. |
Evidence And Verdicts¶
| Verdict | Meaning |
|---|---|
VALID |
Output is consistent/useful enough under the selected policy. |
INVALID |
Output conflicts with consensus, schema, constraints, or policy. |
RETRY |
Caller reruns with a refined task, stronger policy, or different route. |
NEEDS_SPEC |
Task is underspecified or structurally unclear. |
Evidence bundles can include consensus outputs, outlier flags, constraint verdicts, cluster statistics, miner/model metadata, and readable reasons.
Network Alignment¶
| Network phase | Corgent milestone | How they connect |
|---|---|---|
| Testnet phases 1-2 | v0.1 spec and stable testnet agent. | Corgent behaves as a smart router client and starts collecting validation data. |
| Testnet phases 3-4 | v0.2 validation engine and scoring. | Corgent adopts tiered validation and data-driven thresholds. |
| Mainnet preparation | v0.2 to v0.5 transition. | Product policies, payment paths, and service metadata harden before production traffic. |
| Mainnet-Lite | v0.5 default trust oracle in Virtual. | Virtual/ACP/ERC-8004 agents begin using Corgent for higher-trust compute and validation. |
| Mainnet-Full | v1.0 arbitration and trajectories. | Heavy validation and arbitration flows move toward low-cost COR L3 execution. |