The Execution Layer AI Agents Have Been Waiting For: Inside AVM’s Race to Power the Agent Economy

Innovations and Expansion

AVM’s core technical innovation is the integration of the Model Context Protocol as the standard interface between LLMs and code execution. MCP is a lightweight, JSON-based protocol that exposes tools to language models—in AVM’s case, primarily through a run_code interface that accepts agent-generated code, routes it through an MCP server, executes it in isolated containers on AVM nodes, and returns verifiable outputs.

The architecture eliminates the traditional friction of connecting AI reasoning to actual computation. Instead of agents generating code that developers then manually deploy and execute, the agent calls AVM’s tools directly, execution happens automatically in sandboxed environments, and results flow back to the agent for further processing or decision-making.

Beyond one-off sandbox runs, AVM has built a tools and toolkits system that lets developers save working code blocks as reusable functions. Once tested through sandbox executions, code can be saved as a tool with versioning, input parameters, and environment variables. These tools become callable functions in an agent’s permanent toolkit—no need to regenerate the code on every invocation.

Toolkits take this further by grouping multiple tools into structured collections that share environment variables and can execute in sequence. Developers can design domain-specific toolkits for web scraping, data processing, or ETL pipelines where complex workflows are composed from atomic building blocks. Both tools and toolkits can be made public, creating a marketplace dynamic where community-built utilities become available across the entire ecosystem.

The roadmap is structured across five distinct phases running through Q3 2028. Phase 1 foundation work is largely complete with MVP dashboard deployment and core SDK availability. Phase 2 targets 1,000 developers by Q4 2025 with the launch of a sandboxes marketplace where developers can monetize custom toolkits, plus real-time analytics dashboards and multi-modality support for image and video outputs beyond text.

Phase 3 introduces decentralization between Q1-Q2 2026 with self-hosted virtual machines, a proprietary consensus mechanism for coordinating decentralized compute providers, and DAO governance modules. Phase 4 focuses on scaling to 10,000 users through Q3-Q4 2026 with distributed storage for long-term memory, GPU support, and dynamic load balancing. Phase 5 envisions cementing AVM as the global backbone for AI execution with fully equipped virtual machines and AVM Code—described as a Claude Code-style development environment enabling agentic coding loops from zero to goal.

Ecosystem and Utility

The current architecture operates on distributed cloud infrastructure with an alpha dashboard for monitoring compute tasks. Future versions will incorporate what AVM calls a VM Solver that enables peer-to-peer contributions of compute resources, reducing centralization and enhancing network resilience through distributed node operations.

Developer integration happens through two primary pathways: a high-performance HTTP API for direct LLM-to-AVM communication, and comprehensive SDKs including @avm-ai/avm-vercel-ai for TypeScript and @avm-ai/avm-mcp for MCP server implementations. These SDKs support integration with the Vercel AI SDK and MCP-compatible clients like Claude Desktop and Claude Code.

Live use cases documented in the platform span multiple practical applications. Code generation evaluations let developers auto-benchmark LLM output for correctness and safety by executing generated code in secure environments and asserting results against expected outputs. Data analysis workflows enable agents to generate analysis scripts, execute them across AVM nodes with libraries like pandas and matplotlib, and retrieve visualizations or summary statistics. Data extraction capabilities support smart scraping where agents delegate web scraping logic to LLMs, execute parsing code safely through AVM, and obtain structured CSV or JSON outputs. Data transformation pipelines automate format conversions by having LLMs produce transformation functions that execute in AVM’s sandbox and return outputs conforming to specified JSON schemas.

The platform’s flexibility manifests in real workflow examples: validating LLM-generated functions across diverse test cases before production deployment, testing smart contract logic for Solidity compliance, analyzing sales data for business reporting, generating treasury reports for DAOs, scraping product data from e-commerce APIs, extracting cryptocurrency market data from sources like CoinGecko and Dune, and normalizing blockchain wallet data for subgraph indexing.

The economic model centers on a planned token that AVM positions as the coordination layer enabling true decentralization. According to their investor documentation, the token will incentivize global node operators to contribute compute resources, create sustainable demand through pay-per-execution economics, enable decentralized governance of protocol parameters, and fund network expansion through transaction fees. The token is scheduled for launch in Phase 2 during Q4 2025 as part of the expansion toward a permissionless node network targeting 100+ nodes by Q1-Q2 2026.

Bottom Line

AVM is building infrastructure for a computing paradigm that’s emerging right now—autonomous agents that need to execute code independently, securely, and at scale. The positioning is sharp: they’re not competing in the crowded LLM space but rather building the missing execution layer underneath it.

The proof points are tangible. Working MVP with operational Python and TypeScript environments, functional SDK available to beta testers, MCP integration live with major AI platforms, documented use cases spanning evaluation, analysis, extraction, and transformation workflows, and a credit-based system already processing real executions. These aren’t theoretical capabilities—developers can authenticate and start running sandboxed code today.

What makes this potentially sustainable beyond hype cycles is the architectural focus on becoming protocol-level infrastructure rather than an application. By building on MCP as an emerging standard and designing for eventual decentralization with tokenized node incentives, AVM is positioning to become embedded in the agent execution stack rather than competing as a standalone service.

Critical dependencies center on execution speed and timing. The roadmap is aggressive—moving from MVP to 1,000 developers in one quarter, then launching token and decentralization mechanisms within six months after that, and targeting 10,000 users by end of 2026. Success depends on developer adoption accelerating fast enough to justify the infrastructure buildout, MCP gaining broader traction as a standard beyond early adopters, and the transition from centralized cloud infrastructure to decentralized node operations happening without service degradation.

The genuine potential here is real. If AI agents become the dominant interface for how we interact with software—which is the direction the entire industry is moving—then secure, scalable execution infrastructure becomes critical infrastructure. AVM isn’t trying to be the smartest model or the best interface; they’re building the pipes that make agent actions possible. That’s a fundamentally different value proposition, and in infrastructure plays, being early to a standard that wins often matters more than being the most feature-rich. The question isn’t whether agent execution infrastructure will be valuable; it’s whether AVM can build the network effects and technical moats fast enough to own that layer before larger players enter the space.