AgentBuildKit
Build agents. Not demos.
AgentBuildKit is the production layer around your agent framework.
A model alone is not an agent. AgentBuildKit is the harness that turns a model into something you can route, constrain, approve, audit, and evaluate.
Model + Harness = Agent
It currently builds on Microsoft Agent Framework. Each agent declares its own provider and model. OpenAI and Anthropic are the built-in providers, and one workflow can mix them. Deterministic Python owns routing, tool permission, and whether a side effect may run.
What it adds
- Models. Per-agent provider and model selection.
- Agents. Typed inputs and outputs.
- Tools. An allowlist on each agent. Reads can run. Side effects wait.
- Workflows. Deterministic routing across agents.
- Guardrails and human approval. A side effect pauses until a person decides.
- State and audit. SQLite pause and resume, idempotent side effects, and an audit trail that records the provider and model.
- Evaluation. The same run can be checked against expectations.
Quick start
Public install
npx agentbuildkit
Available after the first npm and PyPI release. Neither package has been published yet.
After that release, the command installs the Python CLI with uv and then runs agentkit doctor. In an interactive terminal it offers the offline demo.
Development install
From a clone of this repository, before either package is published:
node npm/agentbuildkit/bin/agentbuildkit.js --source local
That installs this checkout into an isolated uv tool environment. --upgrade reinstalls it. --no-demo skips the demo prompt. If uv is missing, an interactive terminal asks before running Astral's installer. A non-interactive run installs uv only when you pass --install-uv.
To work on the Python package directly:
uv sync
uv run agentkit demo
Python 3.12 is required. uv manages that interpreter. The bootstrapper does not change the system Python.
Commands
agentkit version
agentkit doctor
agentkit demo
agentkit doctor checks the installation. Missing OpenAI or Anthropic credentials are reported and are not a failure.
agentkit demo is offline. It needs no API key and it does not call a model provider. It shows the harness: an agent routes the request, a specialist reads an allowed tool, an unauthorised tool is rejected, a side effect pauses for approval, the run resumes after that decision, a reviewer checks the result, and the evaluation passes or fails from the audit trail.
agentkit demo --approve and agentkit demo --reject make that decision without a prompt.
Architecture
Model
↓
Agent provider and model belong to that agent
↓
Tools allowlist; reads run, side effects wait
↓
Workflow deterministic routing
↓
Guardrails / approval human decision before a side effect
↓
State / audit SQLite pause, resume, and audit events
↓
Evaluation expectations checked against the trace
A workflow can mix providers:
alex → OpenAI / one model
taylor → OpenAI / another model
jordan → Anthropic / another model
The names above are the offline demo's agents. The provider and model are fields on each agent definition. The workflow that receives the clients does not branch on the provider. The offline demo uses scripted clients, so those calls are not OpenAI or Anthropic requests.
The offline demo
The demo is a loading-bay request: book a technician for tomorrow.
- Alex, the orchestrator, chooses the
serviceroute. - Taylor, the service specialist, calls
read_ticket. That read is allowed. cancel_ticketis not on Taylor's allowlist, so it is rejected.schedule_visitis a side effect. The run pauses and stores the approval in SQLite.- A person approves or rejects it. Approval runs the visit once. Rejection does not.
- Jordan reviews the outcome.
- The evaluation checks the trace, including the route, the rejected tool, the approval, and the audit record.
That is the production harness running locally, with the model text supplied by a script.
Current scope
This release includes the Python package agentbuildkit, the agentkit command, and the Node bootstrapper in npm/agentbuildkit. Microsoft Foundry is an optional extra and is not installed by a normal sync.
Skills, persistent agent memory, a dashboard, and Agent Bundles are not in this release.
examples/customer_service is a separate example with fixture data. It is not the offline demo and it is not part of the installed package.
Further detail is in docs/architecture.md.
Development checks
uv run ruff check .
uv run ruff format --check .
uv run pyright
uv run pytest
Bootstrapper tests, from npm/agentbuildkit:
npm test
License
Apache License 2.0. See LICENSE.
Metadata
Release files for agentbuildkit 0.1.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| agentbuildkit-0.1.0.tar.gz | 345.4 kB | With provenance | Details for agentbuildkit-0.1.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| agentbuildkit-0.1.0-py3-none-any.whl | Python 3 | none | any | Details for agentbuildkit-0.1.0-py3-none-any.whl |
Total release size: 479.8 kB
Release files / agentbuildkit-0.1.0.tar.gz
| Download URL | agentbuildkit-0.1.0.tar.gz |
|---|---|
| Size | 345.4 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
b1623f178f53dc206d781fc82cc37680575b09007e68cccba6bbfa41be39c632
|
|
BLAKE2b-256 checksum How to use checksums |
fdf0299c3bc806ef7ee56a9d6836ad9cc2ad574f28672f725817bfe96bb9c51c
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 30, 2026.
Transparency logRelease files / agentbuildkit-0.1.0-py3-none-any.whl
| Download URL | agentbuildkit-0.1.0-py3-none-any.whl |
|---|---|
| Size | 134.4 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
216632b5898858715fbf696bae9ffb237ee94aa429e3c2d598dda6993c34434b
|
|
BLAKE2b-256 checksum How to use checksums |
13f1e356538a2c3cf965e5c169148bc136418392775991ec9a1c3689e27958e4
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 30, 2026.
Transparency logSecurity for agentbuildkit 0.1.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
Both files in this release have been uploaded with provenance metadata.
Attestation for agentbuildkit-0.1.0.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Oct 30, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.