Skip to content

Roadmap

The core API (questions, the engine, decisions, traces) is intended to stay stable; providers, adapters and benchmarks will grow. Plans change with what people build, so open a discussion if something here matters to you, or if something that is not here should be.

Done in 0.3

  • Shadow mode with agreement, projected savings, fitted thresholds and a verdict per question, and an audit mode for live engines.
  • Adapters for LangGraph and the OpenAI Agents SDK, and framework-neutral route and gate helpers (tested with LangChain tools and Pydantic AI).
  • Deadlines and spend limits on the engine, and per-run spend caps.
  • A decision server with a System One compatible endpoint, and MCP tools.
  • Traces to labeling rows, and a drift report between two periods.

Done in 0.4

  • The decision benchmark 1.0: eight public tasks across the four question kinds, frozen rows with published hashes, result files with every prediction, recomputed metrics and open submissions.

Done in 0.5

  • Habits: one line under the OpenAI or Anthropic SDK (or an OpenAI-compatible proxy for any language) that records an agent's LLM calls, finds the decision sites, trains local students from the LLM's own answers, answers a site only after a lower confidence bound on agreement clears the target on past calls and again on a probation of new calls, audits itself and suspends a site whose agreement drops.
  • thinkless habits scan reports which calls in an existing log are decisions and what they cost.

Next

  • Pretrained students on day one. GLiNER, Laya, Jev and any provider as Habits students, gated the same way, so a site can be answered before it has seen hundreds of calls.
  • Stronger students. An optional embedding-based student for sites where hashed n-grams fall short, chosen per site by the same held-out measurement.
  • More of the wire. Local answers for streamed Responses and Messages calls, the Gemini API, and Bedrock.

  • From a site to code. Generate the engine and questions for a proven site, for teams that want rules and hosted decision models in the loop.

  • More models on the decision benchmark. Jev and other hosted decision models, frontier LLMs as references, and Hugging Face classifiers. Version 1.0 has eight tasks and open submissions; later versions add multi-turn safety cases, more languages and harder out-of-scope sets.

  • Cross-request batching in the server. Collect the questions of concurrent requests for a few milliseconds and send them to each local model as one batch, which needs a batch entry point on providers.
  • Calibrated LLM confidence. Several hosted models expose token log probabilities; an LLMDecider that reads them would give LLM answers a real confidence, so they could be thresholded like every other provider.
  • Jev and the direct Anthropic SDK path in the published benchmarks. Both are unit tested (Jev against TypeSafe's own SDK models) but have not been benchmarked live.

Later

  • A trace server with search across runs, for teams that outgrow single HTML files.
  • Learned routing: pick the provider order per question from calibration data instead of by hand.
  • More demos: document processing and an IT operations agent.

Not planned

  • A general agent framework. ThinkLess is a decision plane and a tracer that work inside whatever loop you already have.
  • Hosting models. Providers call models; running them is left to the tools that do it well.