Roadmap & ideas: what's shipped, what's next, what's deliberately not

[dep0we]

Putting the current state of the roadmap in one place so proposals have somewhere to land.

Shipped backend Protocols (the swappable seams)

• MemoryBackend, LLMBackend, JudgeBackend, LockBackend, LogBackend, AgentProfileBackend, ToolRegistryBackend, MandateBackend

Each ships with at least filesystem + one alternate reference impl (SQLite, Redis, etc.), and a Protocol contract for third-party adapters (Postgres, vector stores, hosted SaaS) without forking core.

Next on the protocol roadmap

• PersonaBackend, CorpusBackend, PolicyBackend — filesystem-default-only today; Protocol contracts come next
• Cost-guardrail alert dispatch — currently log-backed only, external channels (Telegram / email / webhook) tracked in #70
• Multi-runtime parity — same agent definition from cron, Claude Code skill, ChatGPT skill, embedded Python (mostly done; ChatGPT-skill runtime is the gap)

What's deliberately not on the roadmap

• A hosted SaaS version. The whole bet is that your agents live in your folder. If someone wants to build a hosted adapter against the existing Protocol seams, great — but the core framework stays operator-owned.
• A web UI for editing agents. The vault is the interface. Your text editor is the UI.
• LangChain / LangGraph compatibility shims. Different shape. The frameworks can coexist; this one doesn't try to be a drop-in.

Proposing something new

Before opening an enhancement issue:

1. Skim the 14 design rules in CLAUDE.md — they're the taste boundaries
2. Post the rough shape here first if you're not sure whether it fits
3. If it touches agent.call(), _capture.py, _costs.py, _locks.py, or a Protocol surface — discuss here or in an issue before opening a PR

What's missing from the roadmap that you wish was on it?