01
Bounded by design
The harness fixes the task, tools, context, turn limits, cost ceiling, and guardrails.
Open source · Apache-2.0 · Python 3.11+
Confine an agent to one job.
Make success provable.
An agent is a model plus its harness. hiveloom lets you—or the coding agent you already use—declare that harness as a folder. A smaller model can then execute the task repeatedly inside a boundary the runtime enforces.
uv tool install hiveloom
what makes the agent
model
general reasoning
+
harness
task · tools · context · loop · validators
=
agent
bounded action
success only when the checks pass
Why hiveloom
More reliable than a prompt.
Narrower than a general agent.
02
Success is a verdict
The model can finish, but only deterministic validators can mark the run successful.
03
The folder is the product
Commit it, deploy it, inspect its runs, and improve it from versioned evidence.
Task confinement, stated precisely.
Builtin file access is rooted, shell commands are allowlisted, and
foreign harnesses are trust-gated. Custom code still runs with the
host process's permissions; use a container or OS sandbox for
untrusted code.
Agent-native by design
Built by an agent.
Run by a smaller one.
Your coding agent can inspect hiveloom's schema and catalog, then create the harness through a JSON-speaking CLI. Every change is validated and rolls back on error—the author does not bypass the boundary it is defining.
- 01Builder agentturns the task into a declared system
- 02Harness foldercaptures tools, limits, policy, and checks
- 03Small executor modelrepeats the bounded task economically
- 04Verified resultreturns to an agent, API, or application
Measured, not assumed
A small model can achieve much more inside the right harness.
On a checked-in article extraction benchmark, the harness helped Claude Haiku hold a contract it otherwise failed. Other models improved less or not at all—which is why hiveloom measures value per task and model.
See the evaluations and caveats →- Task success
- 3%→65%
- Hallucination
- 96%→11%
- Cost per success
- $0.2212→$0.0186
Get started
From task to checked result.
Build explicitly with the CLI or let a strong model generate the same validated artifact. Exploration and dry runs do not call the executor model.
# Define one job
hiveloom init ./summarizer \
--name summarizer \
--task "Summarize a file into JSON."
# Bound its capabilities and acceptance
hiveloom add tool --builtin file_read \
--dir ./summarizer
hiveloom add validator --builtin output_schema \
--schema-file schema.json --dir ./summarizer
# Run with a machine-readable verdict
hiveloom run ./summarizer \
--input notes.txt --jsonOptional development UI
See the boundary and the evidence.
The workbench puts chat beside the exact harness state: spec, runs, traces, version comparisons, and gated improvements.
npx hiveloom-workbench
Tour the workbench →
