Skip to content

Latest commit

 

History

History
35 lines (29 loc) · 3.33 KB

future-features.md

File metadata and controls

35 lines (29 loc) · 3.33 KB

Future Features and Enhancements

Planner and Execution Engine

  • Adaptive planning modes: Toggle between LangGraph-first execution and long-horizon semantic-kernel style planning, with dynamic step ceilings derived from workload metadata.
  • Constraint-aware scheduling: Integrate budget/timebox constraints directly into planning, surfacing soft vs hard limits in traces and UI.
  • Tool routing intelligence: Introduce learned/heuristic scoring for tool selection that blends description similarity, historical success, and latency budgets.
  • Resumable workflows: Add checkpointing for interrupted plans with idempotent step replay and trace stitching.

Memory, RAG, and Knowledge Ops

  • Pluggable vector backends: Ship connectors for popular self-hosted stores with async batching, retention tiers, and deterministic local fallback search.
  • Profile-aware context: Enforce per-profile isolation across ingestion, search, and caching; surface profile provenance in every trace and response.
  • Offline embeddings: Provide multiple deterministic embedding strategies (hashing, TF-IDF, SIF) with benchmark harnesses and guardrails against remote inference.
  • Document lifecycle hooks: Add ingestion/expiry hooks for governance (redaction, PII scrubbing, retention enforcement) plus audit logs.

Registry, Sandbox, and Security

  • Registry hot-swap automation: CLI and CI jobs that validate registry edits, regenerate docs, and roll forward/back with deterministic sorting.
  • Policy packs: Bundled prompts and adapters for prompt-injection defenses, egress controls, and red-team scenarios with kill switches.
  • Sandbox profiles: Granular py/node slim/full profiles with mount templates, read-only defaults, and exec caps; per-tool overrides documented in registry fragments.
  • Access budgets: Built-in budget enforcers (warn|block) with per-profile quotas, rolling counters, and alerting hooks.

Observability & Operations

  • Unified telemetry: Out-of-the-box dashboards for Langfuse/OpenInference/Traceloop spanning registry discovery, planning, execution, and memory connectors.
  • Trace lineage: Cross-component trace_id propagation visualized in UI with drill-downs for tool invocations, retries, and budget events.
  • Health & chaos: Synthetic probes and chaos toggles for registries, sandboxes, and memory backends with SLO-aligned alerts.

Integrations and Surfaces

  • CrewAI/AutoGen adapters: First-class adapters that respect coordinator/worker semantics, shared memory, and deterministic tool import guards.
  • LangServe-ready APIs: FastAPI deployment profile with streaming responses, auth/budget middleware, and rollback-safe config toggles.
  • Workspace UX: Dashboard panels for registry status, budgets, and trace timelines; sandbox diagnostics and tool discovery search.
  • SDK ergonomics: Typed interfaces for planner/worker/coordinator clients, including Python/TypeScript SDKs with generated schemas.

Testing, Tooling, and Release

  • Evaluation harness: Self-play and MCP registry conformance suites with reproducible seeds and coverage gating.
  • Guardrail verifiers: Automated checks for AGENTS.md anchors, registry determinism, and tool import restrictions in CI.
  • Release automation: Tagging pipelines aligned to VERSION.txt, changelog templating, and backport automation for critical fixes.