Shared REPL/eval kernel across concurrent agents: globals get clobbered mid-run

Problem

In a harness where a main agent and several subagents each have an eval / persistent-kernel tool, the kernel namespace can be shared across all of them. During a phase with 5 concurrent agents this bit twice in one session:

A subagent had a global named rows overwritten by a peer mid-run.
The main agent bound p = pathlib.Path(...) in one cell and used p.write_text(...) in a later cell. Between the two cells a peer rebound p to a string. The failure surfaced as:

AttributeError: 'str' object has no attribute 'write_text'

which reads like a type bug in your own code and sends you re-reading logic that is correct.

Why it's hard to spot

The prelude/docs say state persists across cells, which is true -- it just doesn't say whose state.
Short, natural names (p, rows, data, df, res) are exactly what independent agents pick independently, so collision probability is high precisely when concurrency is high.
The error is a plain AttributeError/TypeError at the point of use, arbitrarily far from the clobber. Nothing points at concurrency.
It's nondeterministic: rerun the cell alone and it works.

Mitigations

Prefix globals per agent/task during any parallel phase: _dr_lp, _dr_rows instead of p, rows.
Re-bind in the same cell that uses it. Treat cross-cell globals as unsafe while siblings are live -- cheap to re-derive a Path or re-read a small file.
Prefer passing state through files (local://, artifact paths) rather than kernel globals when agents run concurrently.
If you get an inexplicable type error on a name you're sure you bound correctly, print(type(name)) first and suspect a peer before suspecting your logic.

Generalizes to

Any shared long-lived interpreter session: Jupyter kernels driven by multiple clients, a shared tmux REPL, MCP servers exposing one Python session to several callers, notebook-backed CI runners.