a self-extending agent system. when it hits a task it has no tool for, it writes one,
tests it in a sandbox, and files it away so the next run can just pick it up.
2nd place at the snowflake hackathon.
one pass through the loop. tools only leave the sandbox if they survive the validator.
the problem
an agent is only as good as the tools you hand it. give it a fixed toolbox and it fails
the moment a task falls outside that box. the usual fix is a human noticing the gap and
writing another tool, which puts a person in the loop of every new capability.
how it works
five crewai agents split the job: plan the task, search the registry for an existing
tool, write a new one if nothing fits, test it, and decide whether it is worth keeping.
the registry is a snowflake vector table holding 1536-dimension openai embeddings of
every saved tool. lookup is a similarity search, so the planner finds a tool by describing
what it needs rather than by exact name.
generated code runs in a composio remote workbench, never in the host process.
imports are filtered at the ast level before execution, so a tool cannot reach for
the filesystem or the network unless the task called for it.
a tool is only persisted if it passes validation. failures go back to the codegen
agent with the error attached instead of being written to the registry.
what it means
the toolbox grows on its own. a task that took four steps and a fresh tool on monday takes
one registry lookup on tuesday, and the cost of that capability is paid once instead of
on every run.