Skip to content

ROADMAP.md

Latest commit

 

History

History
79 lines (66 loc) · 4.53 KB

ROADMAP.md

File metadata and controls

79 lines (66 loc) · 4.53 KB
Raw

Roadmap

RiskKernel is the deterministic reliability layer for AI agents — cost/loop/time budgets, crash-resumable runs, human approval gates, tool governance, and an audit trail you own, self-hosted, your keys, no telemetry.

This roadmap is a direction, not a contract — it moves with what users need. The CHANGELOG is the record of what's actually shipped; the public contract and its stability are in api/v1 and COMPATIBILITY.md.

Shipped

The core runtime is built and released:

  • Deterministic budgets — hard per-run cost, token, loop, and wall-clock ceilings that halt a run before it overspends (docs/BUDGETS.md).
  • Crash-resume — kill -9 a run mid-flight and resume it without re-spending (docs/RESUME.md).
  • OpenAI- and Anthropic-compatible proxy — point one env var at RiskKernel and every call (streaming or not) is metered, priced, and budget-enforced (BYO key). Native providers: Anthropic, OpenAI, Ollama (local, key-free), and AWS Bedrock; the long tail is fronted via LiteLLM upstream (docs/PROVIDERS.md).
  • Human-in-the-loop approval — gate side-effecting tools; resolve from the CLI, a local web page, a webhook, or Slack (docs/APPROVALS_SLACK.md).
  • Policy-as-code, enforced per-run — named policy bundles via POST /v1/policies or a reviewed riskkernel.yaml, with a dry-run against recorded runs; a run created under a bundle is governed by its tool allowlist and approval rules, not just its budget (docs/POLICY.md).
  • OpenTelemetry GenAI — export and ingress — emit cost/halt/tool spans into your existing backend (ready-made Grafana + Tempo, SigNoz, and Datadog dashboards), and ingest GenAI spans (POST /v1/traces) to meter apps RiskKernel never proxied (docs/OTLP_INGRESS.md).
  • Spend attribution — roll cost up across runs by team/user/feature (riskkernel audit summary --by metadata.team), with the run name and tags also on the OTel spans so the same grouping works in your backend.
  • MCP gateway — govern an agent's tools/call (allowlist + approval + audit).
  • Compliance evidence export — controls mapped to OWASP / EU AI Act references with a tamper-evident, hash-chained event log (docs/COMPLIANCE.md).
  • Git-native memory — markdown/YAML the agent reads, that you own.
  • Storage — zero-config SQLite by default; an opt-in Postgres backend for multi-instance / HA behind the same Store interface (docs/POSTGRES.md).
  • SDKs — Python (pip install riskkernel) and TypeScript (@riskkernel/sdk), with adapters for the Claude Agent SDK, OpenAI Agents, LangChain, LlamaIndex, CrewAI, AutoGen, and PydanticAI (Python), and the Vercel AI SDK (TypeScript).
  • Operability — a Prometheus /metrics endpoint, a riskkernel doctor setup check, shell completions, a one-command docker compose quickstart, and a Homebrew tap (brew install riskkernel).
  • Measured, low overhead — the enforcement decision is ~150 ns and zero allocations per call, with a reproducible cost benchmark and a timed kill -9 recovery benchmark (exact-once spend) behind the claims (docs/PERFORMANCE.md).

Next

The headline runtime is built out. Near-term work is now driven by what real usage asks for — provider/adapter additions, pricing refreshes, and operability polish as they're requested. Have a need? Open an issue; it shapes this list directly.

Exploring

Bigger bets, gated on demand from real usage:

  • Fleet-level budget throttling — graceful degradation as a fleet of agents approaches a shared budget.
  • Policy replay — re-evaluate a new policy against a corpus of historical runs.
  • Optional semantic memory — an embeddings index, off by default (no vector DB in the core).

Principles that won't change

  • Apache 2.0 core, feature-complete forever — paid is the hosted dashboard and support, never gated runtime features.
  • No telemetry — the OSS binary never phones home (SECURITY.md).
  • Deterministic enforcement — the same state always yields the same allow/deny; no LLM is ever in the enforcement decision.
  • Near-zero adoption friction — meet existing agents where they are; one env var is the gold standard.

Have a need that isn't here? Open an issue — the roadmap follows real usage.