Kirk Stapff
A personal systems notebook

Complex systems.
Clear thinking.

I’m building Workspace, where user prompts and agent turns are nodes in a tree. You can fork from any turn to explore different solutions without poisoning the context. Agents share a working directory and can edit code, run commands and spawn subagents.

Read about Workspace GitHub
WORKSPACECONVERSATIONS AS A TREE
Branch the conversation. Build on the results.An illustrative tree with staggered cards. One prompt branches into research, a failed experiment, and a successful experiment. Research spawns a subagent connected to research.md by a dashed read and update link. The failed experiment continues with a code edit. The successful experiment produces a separate line graph. Select a branch below to highlight its path and outputs. The chart is illustrative, not measured data. PRODUCE GRAPH READ / UPDATE Explore a better modelone prompt / shared goal Researchcompare approaches Experiment Afailed / keep findings Experiment Bsucceeded SUBAGENTReview findingsread files, write notes CONTINUEEdit codeexperiment.py research.md results.svgEXAMPLE OUTPUT CONVERSATIONFILE INTERACTION
FOLLOW A BRANCH

A research subagent reads and updates research.md. The dashed link is a file interaction, not another conversation branch.

Illustrative workflow, not a product screenshot.↑ Try switching the branch

Areas of work

01 — 04 / SYSTEMS & EXPLORATIONS
01 / MARKET SYSTEMS

Custom options matching engines

Matching under constraints, where market structure meets systems engineering.

markets · matching · constraints
03 / BRANCHING AGENT WORKSPACE

Workspace

Conversations as a tree, with branch-scoped history and explicit references. Agents search, edit, run commands, and spawn subagents against one shared project directory.

branching · context · agent tools
Read the post ↗Source on GitHub ↗
04 / AGENT HARNESSES

Code complexity analysis

Designing metrics for agent harnesses to guide coding agents away from producing bloated software.

agents · complexity · metrics

From the notebook

RESEARCH & OPEN QUESTIONS
01
RESEARCH NOTE / MACHINE LEARNING
Recurrent depth in Z1T: a quality–storage trade-off

What happens when a sparse language model applies the same six blocks twice? A controlled study of quality, weight sharing, and unique body storage.

Read the post
02
What belongs in an agent harness?
Essay idea / Agent systems
03
An intuition for energy-based models
Interactive explainer idea / Machine learning
04
Workspace: conversations as a tree
Project note / Branching agents