AI Agent Repos and Starter Templates: Five Places to Begin

Start with the smallest repository that makes your next experiment easy to inspect. These projects provide agent building blocks and official examples; they are not interchangeable application templates or a ranking by GitHub stars.

Last verified September 17, 2026. Rechecked monthly. Download CSV (5 entries).

Start with five

Choose by the hard part of your application: durable state, typed outputs, code execution, workflow coordination, or a coding-agent harness. You do not need all five dependencies.

1. LangGraph · Review release notes before upgrades

   Choose this starting point when checkpoints, branching and human review are central to the workflow. Read the examples before adding graph structure to a simple loop.

2. Pydantic AI · Review release notes before upgrades

   Useful when validated outputs and typed dependencies matter to the surrounding application. Start with one tool and an explicit output type before adding a larger harness.

3. smolagents · Review releases and execution guidance

   A compact place to inspect how a model produces and executes code. Treat its execution and sandbox choices as part of the design, not an implementation detail.

4. OpenAI Agents SDK for Python · Review release notes before upgrades

   Examples cover tools, handoffs and tracing. Useful for learning how these pieces fit together before building a larger workflow around them.

5. Claude Agent SDK for Python · Review SDK and runtime compatibility together

   Start here when the application needs Claude’s agent tooling and permission controls. Inspect allowed tools and hooks before connecting a real workspace.

The directory

The linked maintainer repositories are the primary sources. Read the README, examples, license, release notes and security guidance at the version you intend to use. Inclusion here is not a security audit or a claim that a demo is production-ready.

Make a starter your own

Before connecting customer data, write down the task, allowed tools and stopping condition. Run one official example with disposable inputs. Then replace its example prompt with a small evaluation set drawn from your actual task, including a failure case and a request the agent should decline.

Pin dependencies and retain the example’s license notices. Add timeouts, a cost or turn budget, and a record of model and tool calls. Where tools change external systems, require explicit approval or constrain them to a reversible test environment. A successful run should produce an artifact you can inspect, not merely a confident final message.

When evaluating a framework, test the same task and inputs in each candidate. Record correctness, failure recovery and the work required to understand a trace. A richer feature list is useful only when the application needs it.

Separate the framework from the environment

A Python library can coordinate an agent without providing an isolated machine to execute untrusted code. A hosted runtime can provide that machine without deciding how the agent reasons. The agent development environments guide separates those choices.

For customer-facing engineering work, compare the responsibilities in the FDE and GTM hiring directory.

For the underlying methods, follow the AI engineering blogs. For a reading routine you can import, use the RSS pack.

Change log

1. September 17, 2026

   Added a focused selection with primary-source links, a start-with-five and an export.