Selected work

Built to be
understood.

Projects and experiments, with their design decisions, evidence, and limits kept close to the code.

Agent engineering

pi-go

GoAgent loopsProtocol compatibility

The problem

An agent coordinates model output, tool execution, and session state. Errors at those boundaries can outlive the request that caused them.

The work

The local pi-go project rewrites the upstream pi agent in Go. This site examines its agent loop and extracts a smaller, independently tested teaching implementation. It does not claim authorship of the upstream design.

Design tradeoff

Behavioral compatibility is more useful than an attractive new abstraction when porting an existing protocol. The lab deliberately reduces the protocol surface so its lifecycle is inspectable.

Validation

The accompanying agent-loop lab tests tool results, malformed arguments, model errors, cancellation, and bounded termination using a deterministic model substitute.

Limits

The teaching lab is not the complete pi-go application. Provider compatibility and production behavior require their own validation.

Inspect the evidence

Small teaching implementations; these files do not establish production or GPU performance.

Read the accompanying work

Executable explanations

Engineering Labs

GoPythonCPU PyTorch

The problem

An explanation is difficult to trust when its code, assumptions, and measurements cannot be inspected together.

The work

A companion lab collection built for these articles: bounded admission, an agent loop, structured concurrency, streaming measurements, causal attention, and a cached decoder.

Design tradeoff

Small, deterministic fixtures make correctness visible. CPU runs keep the first iteration reproducible without paid APIs or downloaded model weights.

Validation

Behavioral tests cover failure paths; numerical comparisons check attention and cached decoding. Reproduction commands and raw measurement records accompany the labs.

Limits

Synthetic streams are not model tokens. Tiny random-weight decoders do not demonstrate language quality, GPU throughput, or a production serving stack.

Inspect the evidence

Small teaching implementations; these files do not establish production or GPU performance.

Read the accompanying work