2.works / Projects

Software

Tools and libraries for exploring intelligence and building useful systems.

insideLLMs terminal example comparing baseline and candidate runs and identifying two regressions in dummy-model outputs.
LLM evaluation tooling02

insideLLMs

A Python library and CLI for comparing model behaviour through shared probes, recorded runs and inspectable differences.

GitHub card for dr-gareth-roberts/harness-engineering: “Open-source toolbox for harness engineering: composable primitives for building robust LLM/agent harnesses.”
Agent infrastructure03

Harness engineering

Composable building blocks for the software around an AI agent: context, tools, hooks, memory and telemetry.

Context Engineering Toolkit banner: a stacked context window above a dashed budget line, with the tagline “Pack, debug, red-team, and orchestrate LLM context windows”, 17 packages, and TypeScript plus Python.
Context window tooling04

Context engineering

A toolkit for deciding what goes into an LLM’s context window: pack, score, cache, debug and red-team it.

Intraparietal sulcus simulation with tan, lilac and teal pyramidal banks and a salience topography.
Interactive neuroscience06

Intraparietal Sulcus

A priority-map model of the intraparietal sulcus: 250 pyramidal cells on three banks, and a 14 by 14 salience grid that settles on one locus.

Prefrontal cortex simulation with a layer III population, a cue circle and a coloured spike strip.
Interactive neuroscience07

Inside Prefrontal Cortex

Layer III of dorsolateral prefrontal cortex: 500 pyramidal cells encode a spatial cue, then hold it for four seconds after the cue is withdrawn.

Three pyramidal neurons with gold somata, amber apical trees and a gold membrane trace.
Interactive neuroscience08

Neural Architectures

Three hippocampal pyramidal cells, with a colour legend for soma, apical dendrite, basal dendrite and axon, and a membrane trace of the first cell.

LangChain Command Center with five agent-pattern diagrams and a status of not configured.
Agent pattern console10

LangChain Command Center

Five agent-pattern diagrams on one page. A run calls Claude, and the console stays unconfigured until an API key is supplied.

Mesa-optimizer simulation with twelve agents, a green training platform and a six-step timeline.
Alignment simulation11

Mesa-Optimizer Emergence

Twelve agents chase eight green targets on a training platform, then a second objective turns them magenta and the scene moves to production.

Dorsal anterior cingulate simulation with purple, blue and green cells, a Stroop word and a theta trace.
Interactive neuroscience12

Dorsal Anterior Cingulate

Four hundred neurons of dorsal anterior cingulate cortex, Brodmann area 24, on a Stroop task with a conflict bar and a theta trace.