AgentVM Roadmap

Building in public. Every milestone, every decision, every trade-off — documented here.

Last updated: May 2026

---

Release Timeline

Version	Codename	Target	Status
v0.1.0	Genesis	Q2 2026	✅ Released
v0.2.x	Ignition	Q2 2026	✅ Released (v0.2.3)
v0.3.0	Bridge	Q3 2026	✅ Released
v1.0.0	Launch	Q1 2027	⬜ Planned

---

Phase 1 — Genesis (v0.1.0) ✅

Goal: A working kernel that can spawn, manage, and terminate agent processes with in-memory state.

Motto: "Make it work."

Shipped

• [x] Kernel class — agent registry, process lifecycle, event system
• [x] Agent definition — name, description, config, capability declarations
• [x] Process model — spawn, pause, resume, terminate with state transitions
• [x] Process states: created → starting → running → paused → terminated → crashed
• [x] Event emitter — all lifecycle events fire structured KernelEvent objects
• [x] In-memory process store
• [x] MemoryBus — namespaced key-value store, shared namespace, TTL support
• [x] CLI scaffold: agentvm init, start, spawn, ps, kill, logs
• [x] TypeScript SDK with full type inference
• [x] Unit tests for all core modules
• [x] RFC-001 (Process State Machine Design) — accepted
• [x] Architecture Overview, Getting Started guide, Contributing guide
• [x] GitHub Actions CI/CD

---

Phase 2 — Ignition (v0.2.x) ✅

Goal: Tool routing, message passing, scheduling, and LLM integration turn AgentVM into a real multi-agent runtime.

Motto: "Make it useful."

Shipped in v0.2.0

• [x] ToolRouter — registry, invocation, rate limiting, permission checks
• [x] MessageBroker — pub/sub channels, direct messaging, priority, history, dead-letter handling
• [x] Scheduler — sequential, parallel, race, conditional strategies with dependency resolution and retry
• [x] Kernel.execute() — ExecutionContext with memory, tools, publish, emit, AbortSignal
• [x] Agent tool allowlist enforcement
• [x] RFC-002 (Memory Bus Interface Contract) — accepted
• [x] RFC-003 (Event Schema Specification) — accepted

Shipped in v0.2.1

• [x] AgentContract types — typed input/output, SLA fields (maxLatency, maxCost)
• [x] Kernel.registerTool() and Kernel.createChannel() convenience methods
• [x] MemoryBus.stats, MessageBroker.stats, Scheduler.stats
• [x] Idempotent Process._terminate()
• [x] Scoped rate limit counters (per agent, not global)

Shipped in v0.2.2

• [x] MCPClient — stdio + SSE transports, JSON-RPC 2.0, auto-registers MCP tools
• [x] createLLMAgent() — Anthropic + OpenAI, agentic tool loop, conversation history, token tracking
• [x] createPipeline() — sequential multi-agent pipeline helper
• [x] Built-in tools: http_fetch, json_fetch, shell_exec, file_read, file_write, wait
• [x] registerBuiltins(kernel) convenience function
• [x] Example projects: llm-pipeline.ts, mcp-agent.ts, llm-research-agent.ts
• [x] Tests for LLM agent and MCP client

---

Phase 3 — Bridge (v0.3.0) ✅

Goal: Persistent state, runtime contract enforcement, config system, and framework adapters.

Motto: "Make it connect."

Milestones

M3.1 — Config System ✅

• [x] YAML config file (agentvm.yml) — declare agents, tools, channels declaratively
• [x] Built-in YAML parser (zero dependencies)
• [x] Environment variable overrides via env: section
• [x] Config validation on startup with helpful error messages (ConfigValidationError)
• [ ] Hot-reload for non-breaking config changes
• [ ] agentvm validate CLI command

M3.2 — Persistent Memory Backends ✅

• [x] MemoryBackend interface — stable contract all backends implement (8 methods)
• [x] InMemoryBackend — refactored default, backward compatible
• [x] SqliteBackend — embedded, zero-config, file-based persistence via sql.js (pure WASM)
• [x] MemoryBus accepts pluggable backends via constructor
• [x] Kernel accepts memoryBackend in config
• [ ] Redis backend — distributed cache + pub/sub, connection pooling
• [ ] Migration utility — move memory data between backends

M3.3 — Agent Contract Enforcement ✅

• [x] Runtime input validation against AgentContract.input schema
• [x] Runtime output validation against AgentContract.output schema
• [x] validateSchema() — recursive validator for string/number/boolean/object/array
• [x] ContractValidationError with agent name, phase, and violation details
• [x] SLA enforcement — emits contract:sla:latency event when maxLatency exceeded
• [ ] Pipeline validates type compatibility between chained agents at construction time
• [ ] RFC-004 (Agent Contract Enforcement) — in progress

M3.4 — Resource Tracking ✅

• [x] Surface tokensUsed on ExecutionResult automatically from __llm_usage
• [x] LLM agent propagates __llm_usage to ExecutionResult automatically
• [x] Kernel.stats() — aggregate stats: agents, processes by state, memory, tools, channels, total tokens
• [ ] Per-process token budget — abort execution when limit exceeded
• [ ] Surface cost on ExecutionResult

M3.5 — Checkpointing ✅

• [x] checkpoint(kernel, processId, path) — serialize process metadata + memory to JSON
• [x] restore(kernel, path) — spawn process and restore all memory from checkpoint
• [x] readCheckpoint(path) — inspect checkpoint data without restoring
• [ ] Automatic checkpoint on crash
• [ ] RFC-005 (Checkpointing Strategy) — planned

M3.6 — Framework Adapters ✅

• [x] LangChain.js adapter — tools (toLangChainTools) + memory (toLangChainMemory)
• [x] Vercel AI SDK adapter — tools (toAISDKTools) + usage tracking (createUsageTracker)
• [x] Generic adapter — OpenAI format, Anthropic format, tool executor, MCP server (serveMCP)
• [x] describeTools() debug helper
• [x] Adapter test suite (34 tests)
• [ ] CrewAI adapter (Python-only framework — out of scope for JS SDK)

Phase 3 RFCs

• RFC-004 — Agent Contract Enforcement
• RFC-005 — Checkpointing Strategy
• RFC-006 — Persistent Memory Backend Interface

---

Phase 4 — Launch (v1.0.0)

Goal: Production-grade, distributed, and battle-tested.

Motto: "Make it scale."

Milestones

M4.1 — Distributed Mode

• [ ] Multi-node kernel clusters
• [ ] Process migration between nodes
• [ ] Distributed message broker (NATS integration)
• [ ] Shared state across nodes (CRDTs or distributed locks)
• [ ] Node discovery and health monitoring

M4.2 — Kubernetes Operator

• [ ] Custom Resource Definitions (CRDs) for agents and workflows
• [ ] Auto-scaling based on task queue depth
• [ ] Rolling updates for agent definitions
• [ ] Helm chart for easy deployment

M4.3 — Admin Dashboard

• [ ] Web UI for monitoring all running processes
• [ ] Real-time event stream visualization
• [ ] Resource consumption graphs
• [ ] Tool usage analytics
• [ ] Channel activity monitor
• [ ] Process management (spawn/kill from UI)

M4.4 — Hardening

• [ ] Performance benchmarks (latency, throughput, memory)
• [ ] Load testing at 1000+ concurrent agents
• [ ] Security audit (tool permissions, sandbox escapes, injection)
• [ ] API stability guarantee — semantic versioning commitment
• [ ] Comprehensive documentation site
• [ ] Python SDK

Phase 4 RFCs

• RFC-007 — Distributed Consensus Model
• RFC-008 — Kubernetes CRD Specification
• RFC-009 — Dashboard Architecture
• RFC-010 — v1.0 API Stability Contract

---

Feature Requests & Voting

We use GitHub Discussions for feature requests. The community votes on what gets built next.

How to propose a feature:

1. Open a Discussion in the "Feature Requests" category
2. Describe the problem, proposed solution, and alternatives
3. Community votes with 👍
4. Top-voted features get added to the next phase

How to propose an architectural change:

1. Write an RFC in docs/rfcs/ using the template in RFC-000-TEMPLATE.md
2. Open a PR
3. Community reviews and discusses
4. Maintainers approve or request changes
5. Approved RFCs get scheduled into a phase

---

Build in Public Log

Date	Decision	Context
2026-03-26	Project kickoff	Repository created, roadmap published
2026-03-31	v0.2.0 released	ToolRouter, MessageBroker, Scheduler, Kernel.execute()
2026-04-05	v0.2.1 released	AgentContract types, convenience methods, bug fixes
2026-04-11	v0.2.2 released	MCPClient, createLLMAgent, built-in tools, pipeline helper
2026-05-05	v0.3.0 released	MemoryBackend interface, SQLite backend, contract enforcement, config system, checkpointing, Kernel.stats()
2026-05-05	Framework adapters	LangChain.js, Vercel AI SDK, and generic (OpenAI/Anthropic/MCP) adapters shipped

This log is updated with every major decision. Subscribe to releases to get notified.

---

How to Get Involved

• Build: Pick an open issue and submit a PR
• Design: Write or review RFCs
• Test: Try the latest release and report bugs
• Spread: Star the repo, share on socials, write blog posts
• Discuss: Join GitHub Discussions or Discord

Every contribution matters. Let's build this together.