🧪 Testnet Phase #1 Recap – Integration, Load, and Engine Expansion
Phase #1 was the first structured Testnet week where we focused on one thing: break the pipeline, harden the pipeline, and prove the system scales under real usage across both Testnet-0 (Arbitrum Sepolia) and Testnet-1 (COR #L3).
🔹 Incentivized Manual Testing (E2E + Payments)
- Daily manual testing by community testers across different models and task types
- End-to-end validation: routing → miner execution → validator checks → results
- Exercised both Pay-to-Use (pay per task) and Stake-to-Use (stake-based access / rate-limits)
- General stress-testing across components to surface edge cases early
🔹 Automated Load Testing – 130K+ Tasks Executed
- Automated task jobs ran continuously during Phase #1
- 130,000+ tasks executed, validating stability and throughput under sustained traffic
🔹 Dashboard UI Refinement
- Multiple rounds of UI polish across session/task views
- Improved navigation and usability during high-volume sessions
🔹 #Ollama Engine Support + 65+ Models Live
- Added #Ollama LLM engine support across Testnet-0 and Testnet-1
- 65+ models are now supported (Q1 target: 100+)
- Current limitation is disk footprint: nodes need many models stored to serve on request
- Next up: network-driven model auto-assignment + model-usage-aware routing
- Also added additional embedding models via Ollama to support Validator v2/v3 tuning, since the current embedding stack shows GPU-specific issues on some setups
🔹 SessionQueue Scaling – Read Throughput & Pagination
- Phase #1 load exposed a read bottleneck at high task volumes
- Added missing foundations: precomputed task stats per session + scalable pagination
- Dashboard + Router can now handle large query/read workloads under single sessions
🔹 Corgent Experimental – /delegate + /validate
- First iterations to validate rough E2E behavior and clarify exact requirements
- Established what needs to be strict vs flexible for agent-facing flows
🔹 Router Experiments: #MCP + #x402 + #ERC8004
- #MCP, #x402 and #ERC8004 components are functional inside Router Node today
- Still require refinement and full integration
- These will be re-iterated during Phase #3–#4 to land as router-native primitives
🔹 Roadmap 2026 Docs Ready (Cortensor Network + Corgent)
- Roadmap documentation is now published and aligned for 2026 planning
Phase #1 proved the pipeline can survive both real usage and sustained automated load. Next phases will build on this foundation toward SLA-driven routing, tighter payment flows, and deeper agentic integrations.
https://x.com/cortensor/status/2005225713801277648
https://farcaster.xyz/cortensor/0xc94e762e
Phase #1 was the first structured Testnet week where we focused on one thing: break the pipeline, harden the pipeline, and prove the system scales under real usage across both Testnet-0 (Arbitrum Sepolia) and Testnet-1 (COR #L3).
🔹 Incentivized Manual Testing (E2E + Payments)
- Daily manual testing by community testers across different models and task types
- End-to-end validation: routing → miner execution → validator checks → results
- Exercised both Pay-to-Use (pay per task) and Stake-to-Use (stake-based access / rate-limits)
- General stress-testing across components to surface edge cases early
🔹 Automated Load Testing – 130K+ Tasks Executed
- Automated task jobs ran continuously during Phase #1
- 130,000+ tasks executed, validating stability and throughput under sustained traffic
🔹 Dashboard UI Refinement
- Multiple rounds of UI polish across session/task views
- Improved navigation and usability during high-volume sessions
🔹 #Ollama Engine Support + 65+ Models Live
- Added #Ollama LLM engine support across Testnet-0 and Testnet-1
- 65+ models are now supported (Q1 target: 100+)
- Current limitation is disk footprint: nodes need many models stored to serve on request
- Next up: network-driven model auto-assignment + model-usage-aware routing
- Also added additional embedding models via Ollama to support Validator v2/v3 tuning, since the current embedding stack shows GPU-specific issues on some setups
🔹 SessionQueue Scaling – Read Throughput & Pagination
- Phase #1 load exposed a read bottleneck at high task volumes
- Added missing foundations: precomputed task stats per session + scalable pagination
- Dashboard + Router can now handle large query/read workloads under single sessions
🔹 Corgent Experimental – /delegate + /validate
- First iterations to validate rough E2E behavior and clarify exact requirements
- Established what needs to be strict vs flexible for agent-facing flows
🔹 Router Experiments: #MCP + #x402 + #ERC8004
- #MCP, #x402 and #ERC8004 components are functional inside Router Node today
- Still require refinement and full integration
- These will be re-iterated during Phase #3–#4 to land as router-native primitives
🔹 Roadmap 2026 Docs Ready (Cortensor Network + Corgent)
- Roadmap documentation is now published and aligned for 2026 planning
Phase #1 proved the pipeline can survive both real usage and sustained automated load. Next phases will build on this foundation toward SLA-driven routing, tighter payment flows, and deeper agentic integrations.
https://x.com/cortensor/status/2005225713801277648
https://farcaster.xyz/cortensor/0xc94e762e
❤2👍1👏1
🗓️ 2025 Year-End Wrap - Cortensor
2025 was the year we proved a simple thesis: decentralized inference can be practical if you treat it like infrastructure - not a science project.
We focused on one thing: build the execution loop, validate it under load, and set the foundation for an agent-ready and developer-ready trust layer - so builders can ship reliable AI apps without a high cost barrier.
Minimal E2E MVP is live
We shipped a minimal but real end-to-end pipeline:
Router → Session Queue → Miners → Validators
Payments: Pay-per-use + Stake-to-Use (still early, still needs tuning).
DevNets → Testnet
We completed 6 DevNet phases, kicked off Testnet, and wrapped Testnet Phase #1.
Load testing reached 150k+ tasks executed, validating feasibility under sustained traffic.
Infra & Scaling Bpgrades
Core improvements landed:
- NodePool / NodePoolUtil refactors + data isolation (future SLA routing)
- Session Queue & SessionQueued scaling for high read-throughput + session-heavy workloads
- Dashboard polish for better NodeOps monitoring + rewards/performance visibility
Models + Validation Surfaces Expanded
We rolled out 64+ models (including embedding stacks for Validator v2/v3).
We iterated on /validate (multiple templates) and introduced /delegate as a thin agent-facing SDK surface.
Hackathons pushed the ecosystem forward
We ran 3 hackathons in 2025.
So far, 15+ apps were qualified/rewarded - not just demos, but real feedback loops that pressure-test the network.
Early Agent Rails: #ERC8004 + #x402 + #MCP (experimental)
We ran rough experiments on #ERC8004, #x402, and #MCP to surface blockers early.
Not fully integrated yet - but these are the rails we’ll harden and fold in during 2026.
What Cortensor is becoming
Cortensor is shaping into an execution + trust layer for web3 agents and agentic apps - while staying practical for everyday developers building AI products.
Agents (and apps) delegate work → Cortensor runs redundant inference + proofs → validation/trust signals return → results become usable, reliable, and affordable, without a high-cost barrier to enter the AI app market.
2026 North Star
Build the shared primitives that both agent economies and developers keep needing:
- Execution that scales (reliable inference under load)
- Validation you can verify (PoI/PoUW trust rails)
- Trust signals you can reuse (for routing, SLAs, and agent workflows)
- Low-friction payments (#x402 + agent-native integrations)
That's the path toward becoming the AWS for agent economies - and a builder-friendly execution layer where anyone can ship agentic apps with confidence.
https://x.com/cortensor/status/2006903828713713710
https://farcaster.xyz/cortensor/0x450ea4dc
2025 was the year we proved a simple thesis: decentralized inference can be practical if you treat it like infrastructure - not a science project.
We focused on one thing: build the execution loop, validate it under load, and set the foundation for an agent-ready and developer-ready trust layer - so builders can ship reliable AI apps without a high cost barrier.
Minimal E2E MVP is live
We shipped a minimal but real end-to-end pipeline:
Router → Session Queue → Miners → Validators
Payments: Pay-per-use + Stake-to-Use (still early, still needs tuning).
DevNets → Testnet
We completed 6 DevNet phases, kicked off Testnet, and wrapped Testnet Phase #1.
Load testing reached 150k+ tasks executed, validating feasibility under sustained traffic.
Infra & Scaling Bpgrades
Core improvements landed:
- NodePool / NodePoolUtil refactors + data isolation (future SLA routing)
- Session Queue & SessionQueued scaling for high read-throughput + session-heavy workloads
- Dashboard polish for better NodeOps monitoring + rewards/performance visibility
Models + Validation Surfaces Expanded
We rolled out 64+ models (including embedding stacks for Validator v2/v3).
We iterated on /validate (multiple templates) and introduced /delegate as a thin agent-facing SDK surface.
Hackathons pushed the ecosystem forward
We ran 3 hackathons in 2025.
So far, 15+ apps were qualified/rewarded - not just demos, but real feedback loops that pressure-test the network.
Early Agent Rails: #ERC8004 + #x402 + #MCP (experimental)
We ran rough experiments on #ERC8004, #x402, and #MCP to surface blockers early.
Not fully integrated yet - but these are the rails we’ll harden and fold in during 2026.
What Cortensor is becoming
Cortensor is shaping into an execution + trust layer for web3 agents and agentic apps - while staying practical for everyday developers building AI products.
Agents (and apps) delegate work → Cortensor runs redundant inference + proofs → validation/trust signals return → results become usable, reliable, and affordable, without a high-cost barrier to enter the AI app market.
2026 North Star
Build the shared primitives that both agent economies and developers keep needing:
- Execution that scales (reliable inference under load)
- Validation you can verify (PoI/PoUW trust rails)
- Trust signals you can reuse (for routing, SLAs, and agent workflows)
- Low-friction payments (#x402 + agent-native integrations)
That's the path toward becoming the AWS for agent economies - and a builder-friendly execution layer where anyone can ship agentic apps with confidence.
https://x.com/cortensor/status/2006903828713713710
https://farcaster.xyz/cortensor/0x450ea4dc
🔥4👍2❤1
🧪 Testnet Phase #2 Starts Now
Phase #2 is live - this round is about validation refinement + infrastructure hardening, so Cortensor’s trust layer becomes more deterministic, more redundant, and cheaper to run.
All environments are now active and wired for Phase #2 testing:
- Testnet-0 — #Arbitrum #Sepolia (L2)
- Testnet-1 — $COR #L3 via RaaS
- Testnet-1a — Self-managed $COR #L3 (comparison target)
- Testnet-1b — Self-managed $COR #L3 (internal dev environment)
🔹 Focus – Validation Refinement (PoI + PoUW)
- Tighten Validator v2/v3 scoring so outputs are more deterministic, less “flaky,” and better rubric-aligned.
- Expand redundancy + aggregation testing so trust stays stable under disagreement and partial failures.
🔹 Validator Engine Experiments – #Ollama + Embeddings
- Trial Ollama judge models and multiple embedding variants so validators run reliably across mixed CPU/GPU hardware.
- Reduce validator environment-specific issues while improving scoring consistency and throughput.
🔹 /delegate & /validate Surface Hardening
- Exercise agent-facing flows: policy hints, multi-run delegation, retries/fallbacks, and validation-heavy workloads.
- Confirm response structure stability for downstream apps/agents.
🔹 Infra + Service Reliability
Verify router APIs, indexers, and dashboard feeds stay fresh and consistent under sustained load and recovery scenarios.
Add and test RPC redundancy, failover, and validator backups so the network survives infra hiccups gracefully.
🔹 Gas Optimization & Runtime Tuning
- Profile cortensord + L2 execution paths.
- Optimize SessionQueueValidation, SessionPayment, and validator storage to trim gas + runtime overhead.
🔹 Dashboard – Trust Observability
- Expand real-time PoI/PoUW analytics, validator health views, and endpoint status so operators can actually see trust behavior evolving.
🔹 #L3 Comparison – Self-Managed vs RaaS
- Run apple-to-apple tests on Testnet-1 (RaaS) vs Testnet-1a (self-managed) to measure performance, reliability, and ops tradeoffs.
(Testnet-1b is internal-only for parallel dev validation.)
🔹 Light #MCP Integration (Nice-to-Have)
- Experiment with MCP-style component assignment so the router starts behaving like an agent execution surface (without full protocol rollout yet).
📌 Docs / Plan
- Phase #2 overview & testing focus: https://docs.cortensor.network/community-and-ecosystem/community-testing/testnet-phase-2#testing-focus
- Phase #2 rewards: https://docs.cortensor.network/community-and-ecosystem/community-testing/testnet-phase-2#rewards
Phase #2 is about turning validation into a redundant, cost-efficient, fully instrumented trust layer - the foundation we need before deeper SLA pricing + agentic integrations land in Phase #3 / #4.
https://x.com/cortensor/status/2012497464256033216
https://farcaster.xyz/cortensor/0x1fb7c7c2
Phase #2 is live - this round is about validation refinement + infrastructure hardening, so Cortensor’s trust layer becomes more deterministic, more redundant, and cheaper to run.
All environments are now active and wired for Phase #2 testing:
- Testnet-0 — #Arbitrum #Sepolia (L2)
- Testnet-1 — $COR #L3 via RaaS
- Testnet-1a — Self-managed $COR #L3 (comparison target)
- Testnet-1b — Self-managed $COR #L3 (internal dev environment)
🔹 Focus – Validation Refinement (PoI + PoUW)
- Tighten Validator v2/v3 scoring so outputs are more deterministic, less “flaky,” and better rubric-aligned.
- Expand redundancy + aggregation testing so trust stays stable under disagreement and partial failures.
🔹 Validator Engine Experiments – #Ollama + Embeddings
- Trial Ollama judge models and multiple embedding variants so validators run reliably across mixed CPU/GPU hardware.
- Reduce validator environment-specific issues while improving scoring consistency and throughput.
🔹 /delegate & /validate Surface Hardening
- Exercise agent-facing flows: policy hints, multi-run delegation, retries/fallbacks, and validation-heavy workloads.
- Confirm response structure stability for downstream apps/agents.
🔹 Infra + Service Reliability
Verify router APIs, indexers, and dashboard feeds stay fresh and consistent under sustained load and recovery scenarios.
Add and test RPC redundancy, failover, and validator backups so the network survives infra hiccups gracefully.
🔹 Gas Optimization & Runtime Tuning
- Profile cortensord + L2 execution paths.
- Optimize SessionQueueValidation, SessionPayment, and validator storage to trim gas + runtime overhead.
🔹 Dashboard – Trust Observability
- Expand real-time PoI/PoUW analytics, validator health views, and endpoint status so operators can actually see trust behavior evolving.
🔹 #L3 Comparison – Self-Managed vs RaaS
- Run apple-to-apple tests on Testnet-1 (RaaS) vs Testnet-1a (self-managed) to measure performance, reliability, and ops tradeoffs.
(Testnet-1b is internal-only for parallel dev validation.)
🔹 Light #MCP Integration (Nice-to-Have)
- Experiment with MCP-style component assignment so the router starts behaving like an agent execution surface (without full protocol rollout yet).
📌 Docs / Plan
- Phase #2 overview & testing focus: https://docs.cortensor.network/community-and-ecosystem/community-testing/testnet-phase-2#testing-focus
- Phase #2 rewards: https://docs.cortensor.network/community-and-ecosystem/community-testing/testnet-phase-2#rewards
Phase #2 is about turning validation into a redundant, cost-efficient, fully instrumented trust layer - the foundation we need before deeper SLA pricing + agentic integrations land in Phase #3 / #4.
https://x.com/cortensor/status/2012497464256033216
https://farcaster.xyz/cortensor/0x1fb7c7c2
🔥7❤3🥰2