Build AI agents you can read end to end.
agentkit is a small Python package for building AI agents. Tool calling, a planning loop, memory and evaluation each live in their own short module, so your team can see exactly how the agent works and change it with confidence.
- Language
- Python
- Size
- 14 modules, about 1,300 lines
- Covers
- Tool calling, planning loop, memory, evaluation
- Best for
- Teams who want to understand and own their agent code
- Cost
- Free and open source, always
Why agentkit
Large agent frameworks are quick to start with, but they hide the core loop behind many layers. When an agent picks the wrong tool or loops forever in production, it is hard to see why.
agentkit takes the opposite approach. It is small enough to read in an afternoon. Every step the agent takes is plain Python that you can log, test and change.
- Tool callingPlain Python functions with clear descriptions become tools the model can choose
- Planning loopThe agent plans a step, acts, checks the result and repeats until the task is done
- MemoryKeeps the conversation and working state between steps
- EvaluationTest tasks that check the agent's answers and tool use automatically
How the agent loop works
Every agent built with agentkit follows the same simple cycle.
Receive a task
A user or another system gives the agent a goal, such as "Which tables are stale right now?"
Plan the next step
The model reads the goal, the memory so far and the list of available tools, and decides what to do next.
Call a tool
If it needs information or an action, it calls one of your Python functions with arguments it chooses.
Check the result
The tool's output is added to memory. The agent checks whether it now has enough to answer.
Answer or repeat
It either returns a final answer or goes back to planning, up to a step limit you control.
What it helps your organisation achieve
Prototype on your own tools
Turn existing internal functions and APIs into agent tools in minutes, without adopting a heavy framework.
See why an agent did what it did
Each decision and tool call is a visible step, which makes debugging and reviews straightforward.
Test agents like any other code
Write evaluation tasks once and run them before every change, so behaviour does not quietly get worse.
Teach your team how agents work
The code is short and clear, which makes it a practical way for engineers to learn agent design.
Ways teams can use it
| Use case | Tools the agent calls | Result |
|---|---|---|
| Data platform assistant | Table metadata, freshness and job status lookups | Answers "what is broken and why" in plain language |
| Internal knowledge helper | Search over documents and runbooks | Answers staff questions with links to the source |
| Operations checks | Health checks, log queries, metric lookups | Runs a routine check and writes a short summary |
| Training and workshops | Small example tools | Engineers learn agent design by reading and changing real code |
How to start
Get the code
The source is being prepared for public release on GitHub. Connect on LinkedIn for early access.
-
Write your tools as plain Python functions
A clear name, typed arguments and a one-line description are what the model uses to decide when to call a tool.
def table_freshness(table: str) -> str: """Return how many minutes a table is behind real time.""" ...
Create an agent with your tools and a goal
Give the agent the tools it may use, a short instruction and a step limit.
Add evaluation tasks before you ship
Write a handful of example tasks with expected outcomes, and run them on every change.
Free and open source. Always.
agentkit is a personal R&D project, not a product. There is nothing to buy, now or later.
- No paid tier. Every feature is free for individuals and organisations.
- No sign-up, no licence key. Use it without asking anyone.
- Open code. Read it, change it and run it inside your company.
- Your data stays with you. It runs in your environment. Nothing is sent to LionPact Hub.
Questions about agentkit?
For early access to the code, help applying it to your own tools, or feedback, connect with Rajesh Kaushik on LinkedIn.