Skip to content

ROADMAP.md

Latest commit

 

History

History
163 lines (117 loc) · 7.9 KB

ROADMAP.md

File metadata and controls

163 lines (117 loc) · 7.9 KB

OpenAgentIO Roadmap

Living document. Last updated: 2026-05-24.

Overview

OpenAgentIO is a lightweight, ACP-compatible communication runtime for AI agents. The roadmap is organized into six layers, each with incremental milestones. Breaking changes are batched into major version boundaries; minor versions focus on additive features and stability.

Version Theme Timeline
v0.2 (current) Stabilize Go SDK, HTTP/SSE adapter, middleware chain Now
v0.3 Reliability, protocol state cleanup, first multi-language SDK Q3 2026
v0.4 Advanced orchestration scenes, control plane foundation Q4 2026
v1.0 Production-grade control plane, enterprise features 2027

1. Protocol Layer: Align with A2A Philosophy

Goal: Gradually absorb A2A (Agent-to-Agent) design principles without premature lock-in to a moving standard. EventType semantic overload is the primary technical debt.

v0.3 — Phase Field Introduction

  • Add Phase string to Envelope (schema v2) to carry protocol state (submitted / working / completed / failed).
  • EventType reverts to pure business routing key semantics.
  • Dual-write transition: framework populates both Phase and EventType; consumers read Phase first, fallback to EventType.
  • Update IsTerminal(), HTTP adapter error branching, and SSE event names.
  • Reference: prompts/a2a_prot.md for detailed trade-off analysis (Options A/B/C).

v0.4 — Task Model Evaluation

  • Evaluate introducing a separate Task / TaskUpdate structure for Invoke/Stream paths, decoupled from the generic Envelope.
  • StreamInvoke may return TaskStream instead of Stream.
  • Decision gate: A2A ecosystem maturity + user demand for A2A interoperability.

v1.0 — A2A Interoperability (Conditional)

  • Full A2A protocol bridge if the standard stabilizes and gains traction.
  • Native TaskState enum, OneOf-style response wrappers, artifact streaming.

2. Scene Layer: Richer Orchestration Patterns

Goal: Expand beyond basic Request-Reply, Streaming, and Pub/Sub to cover real-world multi-agent coordination patterns.

v0.3 — Core Patterns

  • Parallel Execution: Invoke multiple SubAgents concurrently, aggregate results (fan-out / fan-in).
  • Agent Handoff: Transfer session context from one agent to another with explicit state passing.
  • Tool Calling: Structured tool invocation with ToolCall / ToolResult event types (reserved in v0.2, enabled in v0.3).

v0.4 — Advanced Patterns

  • Async Task with Callback: Fire-and-forget task creation, completion notification via Pub/Sub or webhook.
  • Circuit Breaker / Degradation: Middleware-level circuit breaker for cascading agent failures.
  • Multi-step Workflow: DAG-style agent chaining with retry policies per edge.

v1.0 — Enterprise Scenes

  • Human-in-the-loop: Pause workflow for human approval, resume via API or event.
  • Multi-tenant Agent Marketplace: Discovery and invocation of third-party agents across tenant boundaries.

3. SDK Layer: Multi-Language Support

Goal: Go remains the canonical reference implementation. Other languages follow after Go API freeze.

v0.2 — Go SDK Stabilization (Current)

  • Freeze pkg/bus public API.
  • Complete golden sample tests as the cross-language specification baseline.
  • Stabilize middleware interfaces (Recover, Trace, Logging, Retry, DeadLetter).

v0.3 — Python SDK Restart

Restart conditions (all must be met):

  1. Go v0.2 released with API freeze announcement.
  2. Core golden sample tests pass stably for 2+ weeks.
  3. schema/envelope.schema.json and pkg/event/golden_test.go accepted as the single source of truth.

Python v0.3 scope:

  • Parity with Go v0.2: Bus (Publish/Subscribe/Invoke/StreamInvoke), in-memory + NATS transports, session context propagation.
  • Middleware chain: Recover, Trace, Logging (Retry/DLQ deferred to v0.4).
  • HTTP/SSE adapter parity optional (v0.4).

v0.4 — Java SDK Evaluation

  • Evaluate JVM ecosystem demand.
  • If pursued, target Kotlin-first with Java interoperability, reusing the same schema and golden samples.

4. Runtime Layer: Reliability & Persistence

Goal: Move from "at-most-once" (NATS Core) to "at-least-once" with persistence, replay, and observability.

v0.3 — JetStream Transport

  • Transport implementation: pkg/transport/jetstream/ with durable streams and consumer groups.
  • Delivery guarantee: At-least-once for Pub/Sub; Invoke/Stream remain at-most-once (idempotency is caller responsibility).
  • Replay: Time-based and sequence-based message replay for audit and debugging.
  • Pull Consumer: Work-queue pattern for backpressure-sensitive consumers.

v0.3 — DLQ & Retry Completion

  • Dead Letter Queue: Exhausted retries automatically republished to {prefix}.dlq.{event_type} with full context.
  • Retry middleware: Backoff policies (fixed, exponential, custom) with acp.retry.attempt metadata stamping.
  • JetStream ack/nack: Integrate with middleware so retries use JetStream redelivery instead of in-process loops where possible.

v0.4 — Observability Integration

  • Prometheus metrics: Messages in/out, latency histograms, error rates per target.
  • Grafana dashboard templates: Agent topology, throughput, latency percentiles.
  • Distributed tracing: W3C Trace Context full support, OTel span linking across agent boundaries.

5. Control Plane: Agent Registry

Goal: Move from static configuration (hard-coded target names) to dynamic service discovery and governance.

v0.4 — Registry Foundation

  • Agent Registry: Central store of agent metadata (id, capabilities, version, health status).
  • Service Discovery: bus.Invoke resolves target name to actual subject/endpoint via registry lookup.
  • Health Checks: Heartbeat mechanism with automatic deregistration on timeout.
  • Dynamic Routing: Route to specific agent instance based on tenant, load, or capability match.

v1.0 — Enterprise Control Plane

  • Credential Distribution: Per-agent TLS certificates or JWT signing keys issued by the control plane.
  • Rate Limiting & Quotas: Per-tenant and per-agent request throttling.
  • Policy Engine: Allow/deny lists for cross-agent invocation, audit logging.
  • Web Dashboard: Visual topology map, real-time traffic flow, drill-down tracing.

6. Monitoring & Dashboard UI

Goal: Operational visibility into the agent mesh.

v0.3 — Metrics & Logs

  • Structured logging (slog) standardized across all packages.
  • Prometheus exporter for key bus metrics.
  • NATS monitoring integration (server stats, connection counts).

v0.4 — Dashboard (Web UI)

  • Topology view: Visual graph of agents, subjects, and invocation edges.
  • Live traffic: Message rate, latency heatmap, error spikes.
  • Trace search: Filter by TraceID, SessionID, or time range.
  • Agent management: Register, deregister, update agent metadata via UI.

v1.0 — Enterprise Observability

  • Alerting rules: PagerDuty/Slack integration for agent health degradation.
  • Cost attribution: Per-tenant message volume and latency attribution.
  • Historical analytics: Trending, capacity planning, anomaly detection.

Cross-Cutting Concerns

Breaking Change Policy

  • Schema version bump (SchemaVersion field) for envelope structural changes.
  • Dual-write periods for field migrations (e.g., Phase + EventType in v0.3).
  • Deprecation notices logged at Warn level one minor version before removal.

A2A Standard Tracking

  • Monitor A2A spec releases quarterly.
  • Maintain an "A2A compatibility scorecard" documenting gaps (tracked in prompts/a2a_prot.md).
  • No commitment to full A2A alignment before v1.0; interoperability bridges are additive, not replacing native protocols.

Python SDK Restart Gate

The Python SDK will not resume development until the three restart conditions in §3 are met. This is a hard gate to avoid chasing a moving Go API.