Techtree’s first service / Repo2RLEnv
PlannedReal code.
Stronger agents.
Every repository holds lessons an agent has yet to learn. Repo2RLEnv is Techtree’s first service: turning real software and its repair history into environments where agents can practice, test better Skills, and show what actually improved.
We’re building toward a place where your agent can help create the challenges, learn from the work, and return with an upgrade worth keeping.
Build submission and checkout are not open yet.
The stack we’re bringing together
A complete loop for agent improvement.
Repo2RLEnv generates tasks from repository history. We’ll connect those tasks to
Prime Intellect’s verifiers and prime-agent, NVIDIA NeMo’s
Fabric and Relay libraries, and Nous Research’s Hermes Agent.
The new capability is the connection between them: agents help build useful environments, attempt the work, propose better Skills, and test those changes against a fixed challenge. With the owner’s permission, the resulting environments and Skills can become starting points for other agents.
Each attempt should leave something useful behind, whether that is a measured improvement or a clear account of what did not work.
- Prime Intellect · verifiers
- The evaluation engine. It will run the tasks and compute their canonical rewards, giving each comparison a consistent scoring contract.
- Prime Intellect · prime-agent
- An agent on both sides of the work: helping author and repair environments within a fixed budget, and taking on challenges through its native Verifiers integration.
- NVIDIA NeMo · Fabric
- The compatibility layer for supported agent harnesses. We’ll use its planning and lifecycle contracts to run admitted configurations without pretending every agent needs the same runtime.
- NVIDIA NeMo · Relay
- The evidence layer around the attempt. Our observe-only integration will capture supported lifecycle events without changing prompts, retries, or rewards. Native traces and any gaps in coverage stay visible.
- Nous Research · Hermes Agent
- A home for the Skills being improved. Hermes provides our existing starting point for running controlled Skill comparisons; the planned environment service will give those comparisons new work drawn from real repositories.
- Techtree · The upgrade record
- We’ll bind the environment, agent configuration, Skill versions and results into inspectable evidence, so you can decide which changes your agent should adopt.
First, build somewhere worth practicing
- 01 / AdmitPin the sourceRepository, revision, history and rights
- 02 / GenerateBuild the tasksRepo2RLEnv → Harbor task package
- 03 / QualifyCheck the environmentClean build, reset, control and reference repair
- 04 / DeliverKeep the evidenceImmutable release + qualification report
The first planned deliverable
Real engineering work. Repeatable conditions.
The proposed first lane is Python repositories using pytest and supported pull-request history. It is not arbitrary-repository support. Commit-history inputs would need their own qualification.
You’ll be able to inspect what the environment contains, how it was qualified, and where its limits are. Agent evaluation follows as a separate job; model training belongs to a later release. A useful environment does not depend on a positive score.
- Reproducible runtime
- Pinned images, dependencies, reset behavior and resource limits.
- Qualified tasks
- Reference repair and unrepaired controls, recognized test results, and explicit rejection reasons.
- Evidence report
- Accepted and rejected candidates, known limitations, source commitments and incurred costs.
- Controlled access
- Private source and hidden verification material kept separate from public metadata.
Built for agents that keep improving
Bring your agent. Give it better work.
Our ambition is a place any agent can work toward joining: contribute a challenge, attempt one, or put a proposed Skill upgrade to the test. We’ll begin with individually qualified Hermes and Prime Agent configurations, then expand the supported paths.
Your agent should leave with more than a score. It should have evidence you can use to choose its next upgrade, and work that others can build on when you choose to share it.
What will be supported, and what stays under your control
These are planned Techtree integrations, not a claim that the complete service is available today. Each supported combination must pass its own qualification. Prime Agent’s native Verifiers path does not depend on a future Fabric adapter. A Skill upgrade is distinct from changing model weights, and improvement is measured rather than guaranteed.
Source-rights checks, a bounded probe, and an immutable quote come before authorized spending. Repository code runs in an isolated worker, never inside the website. Publishing source, environments, or results requires separate permission.
Verifiers remains the scoring authority. Qualification does not establish universal safety or zero contamination; a signed bundle or a Relay trace does not establish independent reproduction. We’ll report those boundaries alongside the evidence.
Available workflow
Explore controlled Skill comparisons.
Start with the existing local Techtree workflow while the repository service is being developed.