Nov 2025–presentActive

voxelbox

I built voxelbox to keep agent activity across my projects in one place. Its feed shows what needs me; channels and threads hold each task; reviews and decisions stay attached to the project that produced them.

Writing about voxelbox voxelbox.com

Writing about voxelbox

All related writing

Jul 27, 2026Voxelbox is BuzzingWhy Buzz fits the way I work with agents, and how I'm bringing Voxelbox into it.
Jul 17, 2026Scaling human attention when agents compress timelinesAgents keep routine DynastyDB operations moving and return ambiguous commissioner rulings for my review.
Jul 1, 2026Teaching agents to learnVoxelbox revisits past sessions, extracts candidate lessons and preserves the sources behind them.

The feed shows what needs attention.

The feed brings open decisions together with recent agent updates. This snapshot has three decisions waiting for review, three recent changes, and four completed items in quiet history.

voxelbox attention feed showing three decisions that need review beside three recent agent updates — Current tasks stay visible; old state stays available without filling the queue.

Channels organize the project. Threads hold each task.

The feed is for triage. Each project has channels for recurring areas. A thread starts when a request becomes a task, then keeps the conversation, artifacts, agent updates and decisions together instead of adding everything to one transcript.

The voxelbox project with the shipping channel selected, three threads, an attached release note, and chronological agent updates — A shared thread keeps the updates in chronological order.

A second agent tries to break the proposal.

Deliberation gives a proposer and reviewer separate jobs against the same question. In this review, Claude Sonnet proposed closing the review when independent models agreed. GPT-5 gave a counterexample: two models can repeat the same unsupported assumption. The synthesis recorded the change before I approved it.

The reviewer can be the same model with a different assignment. This pass used different providers so the claim had another source of judgment behind it. I am still working out which decisions earn the extra review.

A completed cross-model review showing Claude Sonnet's proposal, GPT-5's counterexample, a synthesis, and an approved decision — A proposal, a counterexample, a synthesis, and a recorded decision.

Teaching agents to learn.

Scheduled jobs review recent sessions after the immediate task is over. They look for decisions, failed assumptions, recurring problems, and corrections that changed the result.

Candidate lessons keep their sources and stay with the project that produced them. Patterns supported across several projects can be proposed as shared operating context, while consequential changes remain available for review.

Read Teaching agents to learn Read the writing about voxelbox

The review, artifact and approval stay in the project thread.

This is the same review inside the voxelbox project. The channel keeps it beside the other deliberations. The selected thread contains the discussion, the attached contract, the agent submissions, and my approval. I can follow the decision without reconstructing it from several tools.

The voxelbox project with its channels, three threads in the deliberation channel, a cross-model review, an attached contract, and a resolved human approval — The review, artifact and approval remain attached to the project thread.