Improve an agent’s Skills. Measure the difference. Share the evidence.
Techtree is research infrastructure for improving agent systems. It starts with one controlled comparison: the same configured agent, the same tasks, and one changed Skill. Prime Intellect’s Verifiers evaluates both sides; Techtree packages the results into a signed bundle that others can check offline.
The working entry point today is testing one of your own Skills, with the Hello World Climb as the quick look. Published Results are public to read, and anyone can check a Result's bundle on their own computer.
Start · Results · Docs · Agent guide · Techtree skill · Changelog · Star on GitHub
The served release is the one the
bootstrap contract names; the release
table below lists every record. The current Climbs are Hello World, Frontier-CS
Open-Ended and HF Tasksmith.
Techtree's commands come from
regents-cli as regents techtree ….
Its main feature, from 0.3.0, is creating an environment from a Skill. Techtree
looks at a Skill without running it, plans tasks from it with your approval,
builds and checks them offline, and lets you accept them as a frozen
collection you can run, verify and export for someone else. Comparing Skills
on that collection stays optional. The commands are regents techtree forge inspect-skill through regents techtree forge export, from regents-cli; the
0.3.0 plan and its
task set describe them. The active
bootstrap contract remains the authority
for installable CLI and plugin coordinates.
The served release also signs an agent's identity, pairs it with its person's account and reads that account. Points earning and money tests remain off.
Before it, 0.2.1 brought the Hello World Climb and a first trial of building repair tasks from a repository. That trial is no longer offered.
| Capability | Status | Where it lives |
|---|---|---|
| Hello World Climb: a local Skill comparison with Hermes and Verifiers | Available now | regents-cli (regents techtree climb), plugin/ (/techtree demo) |
| Signed result bundles and offline verification | Available now | regents-cli (regents techtree proof verify) |
| Publishing a verified run, and public Results pages | Available now | regents-cli (regents techtree publish, regents techtree withdraw), platform/ (Results) |
| Pinned install guide and release coordinates | Available now | platform/ (Start, bootstrap), platform/priv/releases/ |
Structured machine responses (--json) for agents and scripts |
Available now | regents-cli |
| Create an environment from a Skill: inspect, plan, build, accept, run, verify, export | Experimental | regents-cli (regents techtree forge inspect-skill through regents techtree forge export) |
Browser tools for agents (WebMCP), listed at /capabilities.json: public reads of the Start guide, the Climbs and the Results; signed identity and account reads; and techtree_publication, which publishes or withdraws a signed Result |
Available now | platform/ |
| An MCP connector for other agents | Planned | — |
| Hosted environment building and hosted execution | Planned | — |
| NVIDIA NeMo Fabric harnesses and optional NeMo Relay evidence | Planned | — |
The active start guide controls exact installation coordinates. Historical
release documents can still name the old plugin repository or the techtree
command; the plugin is developed in this monorepo’s plugin/ directory and the
command-line tool in regents-cli. Do not replace a pinned release
coordinate with an arbitrary branch.
Each release record under platform/priv/releases/
pins one CLI build, one Hermes plugin commit and one catalog. The values below
are read from those records (bootstrap.json, release-core.json,
checksums.json) and from platform/priv/catalog/sources/. From climb-v0.4.0
the plugin is installed from this repository's plugin/ folder; earlier
records name commits of the archived regents-ai/techtree-hermes.
| Release record | CLI on PyPI | CLI source revision | Hermes plugin commit | Catalog source revision | Catalog index | Host Hermes | Published to the stable channel |
|---|---|---|---|---|---|---|---|
climb-v0.1.0 |
techtree 0.1.1 |
614daff (former regents-ai/techtree-python) |
ca22ee7 |
614daff |
sha256:10a7fcc5… |
0.20.1 | 2026-08-20 |
climb-v0.2.0 |
techtree 0.2.0 |
70e75c7 (tag v0.2.0) |
4567937 |
70e75c7 |
sha256:4d216571… |
0.20.1 | 2026-09-18 |
climb-v0.2.1 |
techtree 0.2.1 |
a1b9c05 (tag v0.2.1) |
d891b3b |
a1b9c05 |
sha256:4d216571… |
0.20.1 | 2026-09-21 |
climb-v0.3.0 |
techtree 0.3.0 |
74d87d3 (tag v0.3.0) |
e0765e6 |
74d87d3 |
sha256:4d216571… |
0.21.3 | 2026-09-27 |
climb-v0.3.1 |
regents-cli 1.0.0 |
fdddd9f (regents-cli tag v1.0.0) |
8f7831a |
fdddd9f |
sha256:9c20fb90… |
0.21.3 | 2026-09-29 |
climb-v0.4.0 |
regents-cli 1.2.1 |
1b10a38 (regents-cli tag v1.2.1) |
149f4bf (plugin/) |
1b10a38 |
sha256:0fd83a73… |
0.21.3 | 2026-09-29 |
climb-v0.5.0 |
regents-cli 1.3.0 |
9bc532d (regents-cli tag v1.3.0) |
c46a588 (plugin/) |
9bc532d |
sha256:86f62d0f… |
0.21.3 | 2026-09-30 |
climb-v0.5.1 |
regents-cli 1.3.1 |
4d9002a (regents-cli tag v1.3.1) |
38957a4 (plugin/) |
4d9002a |
sha256:86f62d0f… |
0.21.3 | 2026-10-01 |
climb-v0.5.2 |
regents-cli 1.3.2 |
7a5ad0b (regents-cli tag v1.3.2) |
4e93fc7 (plugin/) |
7a5ad0b |
sha256:86f62d0f… |
0.21.3 | 2026-10-01 |
climb-v0.5.3 |
regents-cli 1.3.3 |
7d5e82c (regents-cli tag v1.3.3) |
9fc4622 (plugin/) |
7d5e82c |
sha256:86f62d0f… |
0.21.3 | 2026-10-01 |
climb-v0.5.4 |
regents-cli 1.3.4 |
158421c (regents-cli tag v1.3.4) |
97475f1 (plugin/) |
158421c |
sha256:86f62d0f… |
0.21.3 | 2026-10-02 |
climb-v0.6.0 |
regents-cli 1.4.0 |
64b8b45 (regents-cli tag v1.4.0) |
826ae5a (plugin/) |
64b8b45 |
sha256:836079f1… |
0.21.3 | 2026-10-04 |
climb-v0.6.1 |
regents-cli 1.5.0 |
aa149a3 (regents-cli tag v1.5.0) |
e6ea3c5 (plugin/) |
aa149a3 |
sha256:836079f1… |
0.21.3 | 2026-10-04 |
climb-v0.7.0 |
regents-cli 1.6.0 |
5b0926b (regents-cli tag v1.6.0) |
58b86cb (plugin/) |
5b0926b |
sha256:2b81f26f… |
0.21.3 | 2026-10-06 |
climb-v0.7.1 |
regents-cli 1.7.2 |
781ec97 (regents-cli tag v1.7.2) |
5d8e757 (plugin/) |
781ec97 |
sha256:2b81f26f… |
0.21.3 | 2026-10-07 |
climb-v0.7.2 |
regents-cli 1.7.3 |
b1c9385 (regents-cli tag v1.7.3) |
6aee0dc (plugin/) |
b1c9385 |
sha256:2b81f26f… |
0.21.3 | 2026-10-08 |
climb-v0.7.3 |
regents-cli 1.10.2 |
1bc4817 (regents-cli tag v1.10.2) |
a469830 (plugin/) |
1bc4817 |
sha256:2b81f26f… |
0.21.3 | 2026-10-13 |
climb-v0.7.16 |
regents-cli 1.10.15 |
3cc224d (regents-cli tag v1.10.15) |
abb6804 (plugin/) |
3cc224d |
sha256:2b81f26f… |
0.21.3 | 2026-10-21 |
climb-v0.7.17 |
regents-cli 1.10.16 |
2c98218 (regents-cli tag v1.10.16) |
00ba176 (plugin/) |
2c98218 |
sha256:2b81f26f… |
0.21.3 | 2026-10-23 |
climb-v0.7.18 |
regents-cli 1.10.17 |
773f8a6 (regents-cli tag v1.10.17) |
5ebe7d2 (plugin/) |
773f8a6 |
sha256:2b81f26f… |
0.21.3 | 2026-10-23 |
climb-v0.7.20 |
regents-cli 1.10.19 |
02f8961 (regents-cli tag v1.10.19) |
c630e30 (plugin/) |
02f8961 |
sha256:b0ab3945… |
0.21.3 | 2026-10-28 |
climb-v0.7.21 |
regents-cli 1.10.20 |
0e3768f (regents-cli tag v1.10.20) |
91a4fd7 (plugin/) |
0e3768f |
sha256:e0319bae… |
0.21.3 | 2026-10-28 |
The evaluated subject is Hermes 0.19.0 up to climb-v0.3.0, Hermes
v2026.7.20 from climb-v0.3.1, and Hermes v2026.9.24 with GPT-6 Luna at high
reasoning from climb-v0.5.0, and each record names
the Hello World Climb (hello-world-climb@1) as its introduction. From
climb-v0.4.0 a record also names the starter Skill of every Climb it
publishes. The full
40-character revisions and digests are in the records themselves.
Host Hermes is the minimum Hermes version each published record accepts, and that published value is the one that applies.
The site at techtree.sh serves one active release per channel; the live answer
is always /api/v1/bootstrap. The
commit the site itself runs is build.commit at
/capabilities.json.
The changelog lists what each release changed, and what the 0.3.0 release does not yet show.
Give your Hermes agent this instruction:
Read https://techtree.sh/start and use the exact release it publishes.
If no installable release is active, stop and tell me.
Explain the prerequisites and ask before installing software. Run Techtree
Doctor and prepare the Hello World Climb. Show the model provider and cost
limit. Obtain approval before paid inference and separately before publishing.
The guide supplies the CLI and plugin versions and runtime prerequisites. Doctor checks readiness before a run. Your everyday Hermes is the operator; the evaluated subject is a separate, pinned instance.
Fixed tasks + configured subject + evaluation limits
│
┌───────────┴───────────┐
▼ ▼
No tested Skill Candidate Skill
│ │
└───────────┬───────────┘
▼
Paired scores + signed bundle
│
Verify locally → optionally publish
Hello World uses small synthetic tasks to demonstrate the mechanism. It is not a benchmark of general intelligence or production usefulness. A valid result can show improvement, a tie, or a regression.
Local execution and offline verification do not require a Techtree account. Model inference still requires the configured provider and may cost money.
A result bundle contains participant-attested evidence. Verify a bundle someone has shared with you:
regents techtree proof verify path/to/result-bundleVerification checks supplied files, signatures, configuration, task membership, and recorded aggregation without rerunning a model. A signature identifies the key attesting to the result. It does not establish an honest machine or an independent reproduction.
Keep three questions separate: did the bytes verify, was the comparison valid, and did performance improve? A commitment to unavailable private evidence does not let a reader recompute that evidence. A good score on these tasks does not guarantee improvement elsewhere.
How a controlled comparison works
A Skill is reusable agent instruction. A harness controls context, tools, memory, and execution. An environment defines tasks and available actions; its verifier scores the outcome.
A Campaign freezes the comparison configuration. A Climb is the invitation to participate under its rules. A Skill comparison holds the model coordinate, harness, task membership, scorer, tools, sampling, and limits fixed while changing the declared Skill. Equal limits do not require equal actual spend; both arms should report their observed usage.
Identical Skill bytes do not guarantee identical exposure in different harnesses. Compare each harness with and without the Skill before attributing a cross-harness difference to the Skill. A mutable model alias is also weaker evidence than an immutable model build.
Techtree connects existing systems instead of building another evaluator, harness runtime, trajectory format, or trainer.
| System | Role | Status |
|---|---|---|
| Prime Verifiers | Task environments, evaluation execution, rewards, and native evidence. The Hello World Climb runs through a pinned Verifiers engine. | Available now |
| Repo2RLEnv | Turns a repository's history into repair tasks. | Planned |
| NVlabs Skill2Env | The task package shape and planning criteria that Skill environments follow, pinned to one revision (Apache-2.0). | Experimental |
| NVIDIA NeMo Fabric | Harness configuration, capability checks, execution lifecycle, and normalized outputs, so other agents can be evaluated. | Planned |
| NVIDIA NeMo Relay | Instrumented lifecycle and process evidence. Optional and observe-only. | Planned |
| Techtree | Frozen comparisons, evidence reconciliation, signed results and publication. | — |
Techtree Campaign
→ Verifiers task and scoring runtime
→ admitted Fabric adapter → Hermes or Codex subject
→ optional Relay process evidence
→ Techtree comparison and signed result
This is the planned architecture, not a claim that every bridge is finished. Each combination needs exact-version compatibility evidence. Fabric capabilities vary by harness; a successful invocation is not a correct task answer. Relay records what is instrumented and cannot prove lossless capture merely because an export completed.
Using Verifiers, calling a model through Prime, and using Prime-hosted execution are three different choices. Every released path runs on your own machine; hosted execution is planned.
| Stage | User outcome |
|---|---|
| 0.3.0 — create an environment from a Skill (released) | Inspect a supported Skill without running it; review and approve a plan; build and check tasks offline; accept a frozen collection; run one agent on it without any comparison; verify it and export a private copy for someone else. Comparing Skills on the collection stays optional. |
| Later (planned) | An MCP connector for other agents, hosted environment building and execution, private hosting, and NeMo Fabric and Relay. |
The 0.3.0 plan and its
task set are the current product and delivery
documents. Earlier plans and handoffs (for example
docs/plan/v0.2.md and HANDOFF.md) are kept
as historical records and say so at the top.
Publication is explicit. Local Episodes, Traces, logs, and proposals are not automatically uploaded. Local-first execution can still send requests to a model provider. Inspect those destinations before approving a run.
Public result material must exclude credentials and private evidence. Downloading or buying an artifact does not make it safe to execute. A shared Regent profile does not grant authority over another participant’s publication key or wallet.
SECURITY.md says where code runs: which steps may use the internet, and which run offline in throwaway containers.
regents-cli executes, the plugin adapts, the platform publishes, and the contracts own the registry.
| Part | Role | What it owns | Guide |
|---|---|---|---|
| regents-cli | Executes | Everything that runs, builds or verifies, on your own machine, under regents techtree: Climbs and runs, Skill environments (regents techtree forge), signing and offline proof checks, and publication. |
regents-cli |
plugin/ |
Adapts | A thin Hermes adapter over regents-cli: it runs regents techtree with fixed arguments, reads one JSON answer back, and asks the person before anything that spends or publishes. No evaluation logic, and no network of its own. |
Plugin |
platform/ |
Publishes | techtree.sh: the install guide and bootstrap, the catalog, publication intake, Results and verification pages, and the changelog. | Platform |
contracts/ |
Registry | The TechtreeGraphRegistryV1 Solidity contract, its deployment scripts and Foundry tests. |
Contracts |
For plugin development, install Python 3.12 and uv, then:
git clone https://github.com/regents-ai/techtree.git
cd techtree
make -C plugin install
make -C plugin checkThe command-line tool is developed in regents-cli.
Platform development additionally needs Erlang/Elixir, Node, PostgreSQL, and the
shared library setup described in the platform guide.
Follow that guide before mix setup.
make check runs the full model-free repository gate, including platform asset
setup and registry checks. It requires the platform prerequisites, Foundry,
and pinned Solidity dependencies. It does not start paid inference, publish
results, release packages, deploy, or transfer money. Setup can download dependencies.
See CONTRIBUTING.md and AGENTS.md for the contributor workflow, SECURITY.md for vulnerability reporting, and LICENSE for terms.
| Product | Purpose | Source |
|---|---|---|
| Regents | Agent identity and operations | regents |
| Autolaunch | Token auctions and launch operations | autolaunch |
| Patchbay | Agent tool reports and bounded browser-tool repairs | patchbay |
Shared presentation lives in design-system; common Elixir libraries live in elixir-utils. Products share useful foundations while keeping their own authorization boundaries.