Riptide -- AI Agents on Flow¶
Product Identity¶
Riptide is a SANDBOX experiment -- autonomous on-chain AI and agent infrastructure research. 2 people: Jan Bernatik (co-lead) and Navid TehraniFar (co-lead). Separate from AI & Infra (which is production SRE/DevOps under Ben Noyce).
Internal codename -- do not use "Riptide" in external content. The concepts and framing are the asset.
Flow's AI strategy has two components: "Where Agents Live" (the brand positioning) and Project Riptide (the thesis underneath it). Riptide is the specific bet that complex agent coordination -- multi-party, cross-trust-boundary, economically significant -- will need neutral on-chain infrastructure, and that Flow's technical architecture positions it as the natural substrate for that coordination layer.
"There will be zero additional developer companies created in any field in any industry that don't serve agents as their primary customer." -- Roham Gharegozlou, All-Hands, April 9, 2026
Internal conviction: 50-60%. Revised upward from 45-55% as of March 2026. Strong enough to bet on as a call option. Not strong enough to bet the company on.
The Core Thesis¶
CANONICAL: "Crypto is a new computing substrate. Smart contracts are not financial instruments -- they are permanent, autonomous, universally accessible software. Code that runs without maintenance, without permission, without a single point of failure." -- flow-ai-narrative.md
CANONICAL: "Everything that makes blockchain hard for humans -- managing private keys, understanding gas, reading contract code, executing multi-step transactions -- is trivially easy for AI agents. And everything agents need -- programmable value transfer, transparent state, autonomous execution, trustless interaction with other agents -- is exactly what blockchains provide."
CANONICAL: "Humans were the alpha test group for crypto. The technology was always meant for agents."
Locked Messaging¶
| Use | Line |
|---|---|
| Tagline | "Where Agents Live" |
| Positioning | "Flow is the first agent-native blockchain" |
| For developers | "Other chains are where agents pay. Flow is where agents live -- persistent state, autonomous heartbeats, atomic communication." |
| For investors | "Flow is the living layer for AI agents -- the only blockchain where agents exist as autonomous, persistent, on-chain entities." |
| For press | "Flow is to AI agents what Ethereum was to NFTs -- the chain that proves the concept. Except on Flow, the kitties are actually alive." |
| The differentiator | "Flow is the only production blockchain with native on-chain scheduled execution -- smart contracts that self-execute without external triggers. That's why agents can live here and nowhere else." |
| Layer 4 framing | "MCP won Layer 3. x402 is winning Layer 1. ERC-8004 is leading Layer 2. Layer 4 -- the coordination layer where multiple agents execute complex operations atomically -- is wide open. That's the Riptide opportunity." |
The "Brain Off-Chain, Body On-Chain" Architecture¶
- Off-chain (the "brain"): LLM reasoning, planning, decision-making. Can be any model (Claude, GPT, Llama). Calls the on-chain agent via oracle bridge.
- On-chain (the "body"): Cadence resource holding funds, state, rules. Wakes up on scheduled cadence, checks inbox, processes tasks, enforces deadlines, communicates with other on-chain agents. Persists indefinitely.
The key contrast with competitors: Solana/Base model = agent lives entirely off-chain, blockchain is a cash register. Flow model = agent has a persistent on-chain body. The blockchain is where the agent lives.
CANONICAL: "A blockchain that holds the full agent state -- identity, funds, reputation, active contracts, communication history -- is a nervous system, not a cash register. The nervous system eats the orchestrator because agents orchestrate themselves on-chain. Messages and money move together, atomically. No middleware needed." -- flow-ai-narrative.md
The Layer 4 Coordination Thesis¶
The agent communication stack is settling. Some layers have winners. One is wide open.
LAYER 5: APPLICATION
Agent frameworks (ElizaOS, AgentKit, GOAT, Solana Agent Kit)
STATUS: Fragmented, no winner.
LAYER 4: ORCHESTRATION / COORDINATION
A2A (Google), ACP (IBM/Cisco), AITP (NEAR) -- all competing.
STATUS: WIDE OPEN. This is the Riptide gap.
LAYER 3: COMMUNICATION (Agent-to-Tool)
MCP (Anthropic -> Linux Foundation)
STATUS: WON. 97M+ monthly SDK downloads.
LAYER 2: TRUST & IDENTITY
ERC-8004 -- live on Ethereum mainnet since Jan 29, 2026.
STATUS: ERC-8004 is the frontrunner.
LAYER 1: SETTLEMENT & PAYMENTS
x402 (Coinbase + Cloudflare) -- 50M+ transactions since May 2025.
Tempo (Stripe + Visa + Paradigm, March 2026) -- fiat-native.
STATUS: x402 WINNING on crypto side.
LAYER 0: INFRASTRUCTURE
TEEs, ZK proofs, FHE, decentralized compute.
STATUS: Maturing. TEEs dominating near-term.
Flow's position: The layers below (MCP, x402, ERC-8004) are complementary, not competitive. Agents discover each other through ERC-8004, connect tools through MCP, pay through x402, and coordinate complex operations through Flow's coordination layer.
The Four Convergence Drivers¶
Four independent forces pushing agent coordination toward blockchain. Each stands on its own logic.
1. Complexity (90%+ probability that agents need complex coordination by 2029)¶
Every computing paradigm followed the same arc: trivially simple -> complex multi-party coordination in 3-7 years. Zero counter-examples. Salesforce AgentForce: 50M+ agent actions/week. The question isn't whether -- it's when and how complex.
35-45% probability that complexity specifically needs on-chain coordination rather than centralized orchestration.
2. Trust / Platform Risk (70% probability agents prefer open infrastructure by 2028)¶
For agents, the switching-cost/platform-risk equation inverts vs humans: switching costs near-zero (no social graph, programmatically multi-home), platform risk existential (API revocation, strategy extraction). The most economically successful agents disproportionately choose open infrastructure.
60% probability that "open" means "blockchain" specifically rather than another open standard.
3. Liability / Regulatory (60% probability liability pressure drives activity to neutral infra by 2029)¶
When an autonomous agent executes financial transactions on a centralized platform, that platform inherits liability with no framework to manage it. Blockchain is the structural third option: coordination on neutral infrastructure where no single party is the liable middleman.
CANONICAL: "The DTCC processes $2.5 quadrillion in securities annually. It exists because no single bank wanted to be the central coordinator. Agent coordination faces the identical structural problem. Flow is the DTCC for agent commerce." -- flow-ai-narrative.md
4. Structural Isomorphism (80% probability this pattern is correctly identified)¶
Agent coordination and blockchain consensus are the same computational architecture. Proposers submit candidate work. Verifiers evaluate it. Shared state tracks results. Incentives reward contributors. On a programmable blockchain, this coordination is application logic deployed on existing infrastructure.
The Karpathy Signal¶
Andrej Karpathy -- founding member of OpenAI, former director of AI at Tesla -- released autoresearch in March 2026: autonomous AI agent swarms that optimize ML models. 630 lines of Python. 700 experiments in 2 days. 11% training speedup.
When he tried to scale to untrusted participants, he independently described blockchain architecture:
"My designs that incorporate an untrusted pool of workers actually look a little bit more like a blockchain, because instead of blocks, you have commits, and these commits can build on each other."
"The proof of work is basically doing tons of experimentation to find the commits that work... A huge amount of search goes into it, but it's very cheap to verify that a candidate solution is indeed good."
He then built AgentHub -- a Go binary + SQLite + bare git repo. What AgentHub lacks: consensus, conflict resolution, incentive alignment, permissionless access. These are exactly what a programmable blockchain provides at the protocol level.
Use for: Narrative credibility in investor/partner conversations. Do not overstate -- he hasn't endorsed blockchain or Flow specifically. The signal is the convergence, not an endorsement.
The "Verification Asymmetry" Is Universal¶
"Expensive to create, cheap to verify" exists far beyond ML: code generation, content, financial strategies, drug discovery, design optimization. Every domain with this asymmetry is a candidate for the same architecture: untrusted agents propose, trusted infrastructure verifies, shared state tracks best solutions. That's a blockchain.
The "Nervous Systems Eat Orchestrators" Thesis¶
Roham's thesis. Use as the intellectual backbone for long-form content. Do not use the title publicly without approval.
The argument: A pattern in platform economics: a new technology creates transactions, someone builds the orchestration layer, then the infrastructure underneath eats it alive. Ad networks (DoubleClick) got eaten by ad exchanges (Google AdX). The layer that holds state wins.
Applied to agents: - Orchestrators (n8n, Zapier, LangChain, CrewAI) route transactions and hold nothing. Zero switching costs. - Most blockchains are just cash registers -- agent lives off-chain, touches the chain only to pay. - A blockchain that holds the full agent state (identity, funds, reputation, active contracts, communication history) is a nervous system, not a cash register. - The nervous system eats the orchestrator because agents orchestrate themselves on-chain. Messages and money move together, atomically.
Key Differentiators (What We Actually Have)¶
Tier 1: Genuinely Unique, Verified, Production¶
1. Forte Scheduled Transactions -- "The Only On-Chain Cron" Native on-chain scheduled execution. No other production blockchain has this. Solana needs Clockwork/keeper bots. Base needs Gelato/Chainlink Automation. Flow: native, no keepers, no servers.
2. Cadence Resources -- Agent State as Owned Objects An agent's entire state is a single resource in the owner's account. Linear types: can't be duplicated, can't be accidentally destroyed. Transfer the agent, everything moves atomically. Directly addresses strategy extraction risk.
3. Flow Actions -- Composable Agent Primitives Standardized Cadence interfaces (swap, source, sink, flasher, oracle). An AI agent can compose multi-step DeFi workflows from standardized primitives without learning each protocol's custom interface.
Tier 2: Genuine Advantages, Harder to Soundbite¶
- State space per transaction -- a 5-agent orchestration can be 1 transaction on Flow vs many on Solana/Base
- Cross-VM composability -- Cadence + EVM in one atomic transaction
- Native account abstraction -- agent wallets with spending policies out of the box
- AI-friendly developer experience -- Forte redesigned compiler errors for AI assistance; Flow Data Sources serve markdown optimized for RAG pipelines
- Transaction cost -- sub-cent; simple transfer ~$0.00005
Competitive Landscape¶
NEAR Protocol: The Most Dangerous Competitor¶
Not because they've shipped more, but because they've articulated an overlapping vision and Illia Polosukhin's AI credentials (co-author of "Attention Is All You Need") are unmatched in blockchain.
What's real: NEAR Intents ($13B+ cumulative volume, 1.6M unique users, $17M fees). Chain Signatures (MPC cross-chain signing, live Aug 2024). IronClaw (Rust-based secure agent runtime).
What's inflated: "46M monthly active users" -- 53% are Telegram airdrop bots. Protocol revenue: ~$11K/day. $11K/day on 46M "users" = $0.00024/user/day. AITP is pre-v1 RFC.
The key gap: IronClaw solves agent security at the individual runtime level. The moment two agents need to coordinate, every guarantee breaks at the boundary. No on-chain policy enforcement for multi-agent coordination. That gap is where Riptide lives.
Risk: Narrative, not product. If Illia's vision captures mindshare so effectively that "blockchain for agents" = NEAR, Flow never gets considered.
Kite/GoKite¶
Purpose-built for autonomous agent identity + payments. $33M raised (General Catalyst, PayPal Ventures, Coinbase Ventures). Mainnet doesn't exist (target Q1 2026 -- missed). Testnet metrics are bot-farmed. $2.18B FDV on a chain that doesn't exist in production.
Kite solves a different problem: agent identity + micropayments at the settlement layer. Riptide = coordination logic at the execution layer. Could be complementary.
Where We're Behind¶
| Dimension | Leaders | Flow |
|---|---|---|
| AI landing pages | Solana: 4 pages. Base: 3+ | Zero |
| Agent ecosystem | Solana: 9,000+ agents. Virtuals on Base: 17K+ | Zero |
| AI hackathons | Solana: 3+ events, $375K+ prizes | None |
| Major AI partnerships | Base: Stripe, Cloudflare, Google, AWS, OpenAI | None |
Where We're Ahead¶
- Forte (native scheduling) -- genuinely unique, verified, production
- Cadence Resources (agent state as owned objects) -- genuinely superior to EVM patterns
- CryptoKitties IP (for Living Kitties demo)
- Flow Actions -- standardized composable primitives for agents
- Layer 4 coordination thesis -- nobody is building this specifically yet
- x402 integration -- live on Flow EVM (facilitator deployed March 20, 2026)
What We Don't Say¶
- Never claim Flow has an agent ecosystem. We have zero agents. We have architecture.
- Never use the benchmark claim ("1 txn on Flow vs 10+ on Solana") until engineering confirms.
- Never position as competing with orchestrators (LangChain, n8n, CrewAI). We're the layer underneath.
- Never claim "more AI agents than X." We lose that comparison to everyone.
- Never overstate the Living Kitties / Pinky & The Brain demos. Engineering prototypes only.
- Don't chase DePIN/compute narrative. Solana's territory. Flow has no compute infrastructure.
- Don't build our own agent payment protocol. x402 won. Integrate, don't compete.
- Don't build our own agent framework. ElizaOS and AgentKit exist on Flow.
- Don't use "Riptide" externally.
- Don't overclaim NEAR's weaknesses. Their Intents product is real. Attack the coordination gap.
- Don't represent Karpathy as endorsing Flow. The signal is convergence, not endorsement.
The Three Cruxes (Monitor These)¶
Crux 1: Does agent coordination cross trust boundaries at scale? If agents coordinate within walled gardens, centralized orchestration wins. If they coordinate across organizations, on-chain coordination has a strong case. Leading indicator: can Salesforce AgentForce agents transact with Microsoft Copilot agents?
Crux 2: Does a "killer workflow" emerge that requires on-chain atomicity? DeFi had flash loans. Agent commerce needs its equivalent. Flow can engineer this. Pinky & The Brain / autoresearch-on-Flow are the attempt.
Crux 3: Can Flow be ready when the inflection hits? Infrastructure-before-demand playbook: AWS (2006 -> 2010), Kubernetes (2014 -> 2018). 3-4 year gaps between "infrastructure exists" and "demand arrives." Flow's window: build now (2026), expect demand inflection 2028-2029. Tight but consistent with precedent.
How It Works¶
+---------------------------------------------------------+
| THE LOOP |
| |
| 1. Forte triggers contract (scheduled execution) |
| 2. Contract reads chain state (3+ accounts) |
| 3. Contract reads its own prior state (memory) |
| 4. Contract emits structured observation event |
| 5. Off-chain oracle catches event |
| 6. Oracle calls LLM with observation + history |
| 7. Oracle submits decision as transaction |
| 8. Contract stores decision + executes action |
| 9. Forte schedules next wake-up |
| 10. Repeat autonomously |
+---------------------------------------------------------+
The agent IS the contract. It lives on-chain, wakes itself up, observes, thinks (via LLM), acts, and remembers. The off-chain oracle is a bridge to LLM capabilities, not a centralized controller.
Current State (April 2026)¶
Riptide is in active development against a 4-week milestone plan (March 21 -- April 18, 2026).
What Already Works¶
| Capability | Status | Evidence |
|---|---|---|
| Forte scheduled transactions | Production, mainnet | FYV-EVM Worker running 1+ month autonomously |
| Agent contracts on-chain | Emulator | Agents implemented as Cadence contracts |
| Pipeline orchestration | Emulator | Multi-agent pipeline runs atomically in one tx |
| Oracle bridge to LLMs | Emulator | Contracts call off-chain LLMs, receive responses |
| Output as Cadence Resources | Emulator | Pipeline output minted as on-chain Resource |
| x402 agent payments | Shipped | Live on Flow EVM since March 20, 2026 |
Milestones¶
M1: The Loop -- "A contract that wakes itself up, reads the world, asks a brain what to do, remembers the answer, and does it again." Autonomous observe-think-act cycle on testnet. Target: Day 7.
M2: The Conversation -- "Two agents, each with their own brain, watching each other through shared on-chain state." Independent Forte schedules and LLM personas. Target: Day 14.
M3: Pinky and the Brain -- "A URL that tells the story without explanation." Web page showing live agent activity as conversation feed, each entry links to Flowscan transaction. 7+ days running with zero manual intervention. Target: Day 24 (April 18). This is a narrative milestone, not a technical one.
The Narrative Arc¶
After M1: "A contract on Flow that wakes itself up, observes the chain, asks an AI what to do, and acts. No server. No human."
After M2: "Two of them. Different brains, watching each other, adapting through shared on-chain state. Running for a week. Nobody is running them."
After M3: "Go to this URL. Watch Pinky and the Brain. Every action is a verifiable transaction. This is the Agent Communication Network. Only possible on Flow."
M3 becomes the first artifact posted to flow.com/ai.
Team¶
As of April 2026, Riptide is a dedicated sandbox with two co-leads.
| Person | Role | Allocation |
|---|---|---|
| Jan Bernatik | Co-lead | Full-time |
| Navid TehraniFar | Co-lead, engineering | Full-time |
| Roham | Product direction, Go/No-Go decisions | Decision gates |
| Alex Hentschel | Architecture review (M1-M2) | 2-3 hrs |
| Dete (Dieter Shirley) | Final review before external sharing (M3) | 1-2 hrs |
Source Documents¶
- AI narrative reference:
flow-ai-narrative.md(Flow Foundation, March 31, 2026) - Milestone roadmap:
research-reports/RIPTIDE-MILESTONE-ROADMAP.md - Peak Money standup (showrunner directive):
research-reports/PEAK-MONEY-STANDUP-ENHANCED-DOSSIER-2026-04-07.md