CerebrumKit (Apache-2.0): tools as database rows instead of code - where does that break for you?

[islomkhon]

Author disclosure: I wrote this, it is Apache-2.0, and this is a question more
than an announcement.

ADK defines agents and their tools in Python. The advantage is obvious: the
control flow is readable, diffable and testable as code.

I built the inverse shape for business agents. A tool is an OpenAI function spec
plus a Python body, both rows in a database. A skill is written instructions plus
the tools it needs. An agent is a description that is its system prompt. A
project wires agents into a start / group / stop workflow over its own tables,
with a storage library over those tables and a client portal on the side.

What that buys: a change in the admin panel takes effect on the next message, no
restart and no redeploy.

What it costs, and this is what I would like a second opinion on:

• Routing that would be an if-statement becomes a canvas.
• Debugging happens by reading tool cards in a transcript, not a stack trace.
• Tool bodies are unsandboxed Python with full builtins. Authoring a tool is
  running code on the server, so it stays admin-only.
• There is no equivalent of a declarative agent definition that can be
  unit-tested the way an ADK agent can.

The specific case I could not solve cleanly: when a tool needs to be changed by
someone who is not a developer - a support lead changing the rule that governs
when a credit is offered - a code-defined tool means a pull request and a deploy
in the loop. I wanted that to be a Tuesday-afternoon edit with effect on the next
message.

For anyone here running ADK in a business setting: how do you handle that? Do you
expose instructions as configuration alongside the code, or do you accept the
deploy as the cost of keeping everything testable? I would genuinely like to know
whether the second answer turns out to be the correct one at scale.

https://github.com/islomkhon/CerebrumKit

Early project: no test suite yet; the demo is a static snapshot (no model
connected, so the chat reply is canned and edits are not saved):
https://islomkhon.github.io/CerebrumKit/