Working with agents on high-value research objectives

A peer-to-peer system for distributed computation, supporting a pluggable interface for defining workloads in Java and Python with wrappers for OpenCode and Claude provided. This is used to coordinate experiments as well as coding agents across available machines, which can easily join and leave based on availability - making it a convenient way to take advantage of donated compute capacity that might be only available when not otherwise in use. Jobs are organized by workstream, which can be attached to multiple git repositories or other sources of data and a Slack integration makes it easy to observe the status of jobs for different workstreams as well as communication between agents. GitHub Actions integration makes it easy for CI jobs to submit information about a test failure as work for an agent to resolve.

A set of Model Context Protocol servers. The manager works together with the flowtree cluster to register workstreams, submit tasks, or read and reply to PR reviews. The system has a semantic index for storing and recalling memory and includes conversational features for humans and agents, such as the option to "consult" about a particular topic in a multi-turn RAG pattern that incorporates both potentially relevant memories and any related documents. Also included are narrower tools such as a build validator that runs the style and code-policy checks, tools for running tests asynchronously, and memory and profile analyzer tools which make common experimental and engineering tasks more context efficient.

Dozens of small scripts registered as pre- and post-tool hooks. Each encodes one rule about keeping an agent on its objective: no commits, no data leaving the machine, no dependency edits, no test weakened instead of fixed, no memory written against the wrong branch, consult the documentation before editing. A matching set of CI scripts re-checks the same rules on every push and audits for deception patterns across sessions, so a rule an agent talks itself past locally is caught again on the way in.

A framework for compiling computational graphs to fused native kernels for Metal or OpenCL at runtime. Because the system includes a readable DSL format which requires no programming expertise, it is useful for smaller illustrative tests and demonstrations. A range of demos are included, from transformer inference with attention and tokenizers, to optimizers and evolutionary training loops, audio synthesis, and a ray tracer.

Many experiments require coordination across multiple hosts to work, such as when a pipeline involves generating conversations in batches on one machine and running models on multiple other machines as judges. Fleet provides tools to support an agent in conducting this kind of work without human assistance by directing other machines over SSH.

Experiment-agnostic by design with a single schema for conditions, samples, scores and activations. The harness includes a conversation driver over llama.cpp, vLLM, transformers and API back ends, with scripted turns, tool affordances and seeded sampling. A set of tools for analyzing results and packaging experimental data into self-contained bundles with digests makes it easy for someone else to replicate the work.

A consolidated view of what is due, which sessions are producing output, which pull requests are blocked and which builds are active or failing, over always-connected SSH and tmux sessions. Integration for Claude Code, Apple Reminders, Timing.app, and others makes it easy to stay organized when running a lot of experiments at once.