Skip to content

roadmap.md

Latest commit

 

History

History
135 lines (105 loc) · 6.21 KB

roadmap.md

File metadata and controls

135 lines (105 loc) · 6.21 KB
Raw

Roadmap

Living roadmap for the WasmAgent organization. Ticked items ship as public repositories under github.com/WasmAgent; unticked items are planned or in progress. This document mirrors gh repo list WasmAgent --visibility public, and its machine-readable counterpart is project-index.json.

Status legend

  • ✅ Public repository exists and is the source of truth for its layer.
  • 🚧 In progress — spec or reference implementation landing.
  • 📋 Planned — not yet public.

Focus & lifecycle

WasmAgent's north star is open infrastructure for provable AI agents. Repositories are ranked by how close they sit to that mission, orthogonal to the maturity tiers in repository-boundaries.md:

  • ⭐ Core (wasmagent-js, wasmagent-protocol, symkernel, agentbom, and the planned wasmagent-py) — runtime, protocol, verification, and trust artifacts. Sustained long-term investment. AEP is their connective standard, sedimented from the shipping runtime and now versioned in wasmagent-protocol.
  • 🛠 Official tooling — CLI, devtools, examples, starters. Landing: @wasmagent/agentbom-cli (in agentbom) and agent-golden-path (the runnable provable-stack reference app) are public; more devtools planned.
  • 🔌 Evidence surfaces (wasmagent-proxy, trace-pipeline, wasmagent-train-replay) — extend Core to specific surfaces. Once their roadmaps land they move to community maintenance; not retired — code and evidence schemas stay put.
  • 🧪 Research (fresharena) — grounds the platform in measured results.
  • 📦 Product, reference & hub (open-agent-audit, bscode, .github) — the commercial audit surface, the reference workload, and this portal.
  • 🎮 Apps — playground, desktop, editor extensions, demos. Planned; no public repos yet.

The machine-readable focus field in project-index.json is the source of truth for this axis.

Layers

Runtime ✅

  • ✅ wasmagent-js v1.x — embedded agent runtime: WASM kernels (QuickJS, Pyodide, Wasmtime, Remote), MCP gateway, MCP attestation, AEP emitter, capability manifests; adapters for A2A, AG-UI, AI SDK, Claude Agent SDK.
  • 📋 wasmagent-py — Python agent runtime sibling; shares AEP schema, Criterion/ConstraintIR protocol, and symkernel HTTP adapter.

Gateway 🚧

  • 🚧 wasmagent-proxy — Proxy-Wasm (Rust) evidence engine for Envoy, Istio, Kong, and Consul: intercepts Agent/MCP/A2A HTTP traffic, applies validation → delta → full recording policy, emits Ed25519-signed AEP records (DSSE envelope). Joins mcp-firewall via shared trace_id.

Verification 🚧

  • 🚧 symkernel — Go symbolic verification backend: cel-go lightweight rules, wazero Wasm sandbox hard-isolation, Z3 SMT satisfiability proofs. OPA-style HTTP service; consumed by wasmagent-js (CelGoVerifier / Z3Verifier) and future wasmagent-py via a thin adapter. Now imports OPA Rego and AWS Cedar policies (translated to CEL, fail-closed — internal/policyimport), positioning symkernel as a Z3-proof backend for existing policy languages. Deploys on Cloudflare Containers.

Workloads ✅

  • ✅ bscode — coding-agent workload on Cloudflare Workers with AEP evidence export, deny capabilities, output taint labels, and RolloutProvenance.
  • ✅ agent-golden-path — ERP-domain (procurement) reference workload: a runnable copilot that mirrors the role bscode plays for coding tasks and proves the full provable-agent chain end to end. Supersedes the planned erp-agent (RFC-0001).

Evidence pipelines ✅

  • ✅ trace-pipeline (evomerge on PyPI) — eval_trust paired statistics, AgentTrustScore stable JSON schema, training-data admission gate; schema compatible with wasmagent-js v1.x AEP.
  • 🚧 wasmagent-train-replay — causal evidence layer for distributed GPU training: reads PyTorch Flight Recorder dumps, builds cross-rank PROV-DM provenance graphs, records Ed25519-signed EpochEvidenceBundles, supports tensor-origin tracing and deterministic replay.

Trust artifacts ✅

  • ✅ agentbom — AgentBOM, MCP Posture, and Trust Passport tooling: @wasmagent/agentbom-core (validator, compliance checker, BOM diff/drift) and @wasmagent/agentbom-cli. Successor to agent-trust-infra (archived 2026-07-29); the library packages and CLI now live here. Trust Passport / AgentBOM canonical schema → wasmagent-protocol.
  • 🚧 EU AI Act Annex IV compliance mapping — draft covers 20/29 Annex IV sub-items; deadline 2026-08-02 (Article 11 effective for high-risk AI).

Audit ✅

  • ✅ open-agent-audit — enterprise audit product with AEP v0.3 adapter; deployed at trustavo.com. @openagentaudit/core, adapters, schema published on npm.
  • 🚧 Regulatory profiles: OWASP Agentic Top 10, NIST AI RMF, ISO/IEC 42001, EU AI Act Annex IV.

Evaluation ✅

  • ✅ fresharena — dynamic, verifiable, adversarial evaluation protocol for coding agents; FAEP schema, submit-then-test adversarial testing, Public Immunity Pool. Results feed back into the evidence chain.
  • 🚧 Technical paper (FreshArena: FAEP protocol + empirical results on JSON Transform World) in preparation.

Project home ✅

  • ✅ .github — org portal; profile, public ledgers, shared org docs, RFC registry, and the source of truth for the full roadmap and project list.

Ops ✅

  • ✅ wasmagent-ops — internal operations hub: generators and ops tooling for the org. Ships no public product; listed for inventory completeness.

Cross-cutting (planned)

  • 🚧 Ingestion adapters: OpenTelemetry GenAI (field-level AEP↔OTel GenAI mapping landed in wasmagent-protocol/docs/AEP-OTEL-MAPPING.md; adapter code pending), Langfuse, LangSmith.
  • 📋 wasmagent-py Python runtime and symkernel adapter.
  • 📋 Cross-repo coherence patrol: org repo list ↔ profile README ↔ cross-repo URLs ↔ roadmap completion.
  • 📋 Trust Passport product module: issuance, verification, renewal (Trustavo).
  • 📋 symkernel Phase 1: wazero sandbox (/v1/sandbox/run), Z3 integration (/v1/verify/z3), generate-error-repair loop.