Improve any agent's Harness / Environment / Skills to use less tokens, perform evals stronger or create better traces Improve any agent's Harness / Environment / Skills to use less tokens perform evals stronger create better traces

Turn know-how and real work into tasks agents can practice on, then see what actually helps them.

• Skill2Env

  Experimental

  What it does. Turns a Skill, the know-how written down for an agent, into a set of practice tasks with checks. Techtree reads the Skill without running it, plans tasks with you, builds and checks them on your computer, and you accept the final set.

  Why. The task set is useful on its own: you can keep it, export it, or hand it to someone else to run, whether or not you go on to compare anything.

  When to use it. You have a Skill and want to know whether agents follow it, or you want a fixed set of tasks to measure each change to it.

  Test a Skill → NVIDIA’s Skill2Env →

• Repo2RLEnv

  Planned

  What it does. Turns a real repository and its history of fixes into practice environments, each one a past problem with a known good answer.

  Why. Real bugs that were really fixed make realistic practice, and their own tests say whether an agent got it right.

  When to use it. Your know-how lives in a codebase rather than a written Skill, or you want agents to practice on your own software’s real problems.

  About Repo2RLEnv →

• Tasksmith

  What it does. The HF ML Tasksmith collection packages real changes to Hugging Face’s machine-learning libraries as tasks. Techtree’s Tasksmith Climb uses six of them to ask whether adding the Tasksmith Skill solves more.

  Why. It is a ready-made test on serious code, so you can see a Skill comparison work without building tasks of your own first.

  When to use it. You want to try Techtree on realistic machine-learning engineering, or measure your own Skill on the same six tasks.

  The Tasksmith Climb → The collection on Hugging Face →

• FrontierSmith

  Experimental

  What it does. Writes open-ended programming problems, each with its own checker that scores an answer from 0 to 1 rather than simply right or wrong. Techtree’s Frontier-CS Climb uses ten problems made with it.

  Why. Open-ended problems have no single answer to memorise, so a better score means better work, and partial credit shows small gains a pass-or-fail test would miss.

  When to use it. Your Skill is about solving hard problems well rather than following steps, and you want to see whether it raises the quality of the answers.

  The Frontier-CS Climb → FrontierSmith →

• Frontier-CS Open-Ended

  Ten open-ended optimisation problems with no known best answer, from Frontier-CS (MIT licence) as chosen by the FrontierSmith authors. With thanks to both.

• Techtree Hello World

  A toy introductory demonstration of the mechanism, not a measure of broad capability.

• HF Tasksmith

  Six real changes from Hugging Face's trl, transformers, diffusers, peft and accelerate repositories, from the HF ML Tasksmith collection. With thanks to its authors. Not a measure of broad capability.