the baro blog
Field notes on parallel coding agents — execution modes, benchmarks, real runs, and the bugs that taught baro to work on its own.
Not Every Task Deserves Parallel Agents
I built baro for parallel agent runs because my tasks decompose that way. Then real users arrived, and their tasks didn't. The bugs that taught baro to pick focused, sequential, or parallel on its own — with both proof runs as open PRs.
Agents That Remember: Shared Memory for Autonomous Agent Teams
Autonomous agent teams need what human teams have — shared memory and real-time communication. baro now ships cross-run memory: every run's decisions persist, and the next run recalls them. Here's what we built, and where it's going on Mozaik.
How We Run a Fleet of AI Agents on One Repo Without Chaos
The hard part of parallel AI coding isn't the agents — it's keeping a dozen of them from destroying each other's work on the same repo. Git worktrees, a story DAG, merge-back conflict resolution, and two production bugs that taught us the difference between 'works on my machine' and 'works in the cloud'.
I Stopped Babysitting AI Coding Agents. I Gave Them a Cloud and a Goal Instead.
One agent in a chat is the wrong shape for real work. Here is what happened when I handed a goal to a fleet of coding agents running in the cloud — and how it differs from Claude Code's /goal.
The Economics of Parallel Coding Agents
I benchmarked Claude Code, baro hybrid, pure-Claude baro, and DeepSeek execution on the same real NestJS feature. The fastest successful run cost four cents.
I tested Claude Code vs OpenAI Codex in my parallel agent setup. Then I built a hybrid.
Three tasks, two CLI backends, $100/mo subscription each. Codex burned a fraction of the budget per run but reached for tools 3–4× as often. Anthropic is killing claude -p subscription billing on June 15, and the calculus shifts entirely. Here's how baro now routes Claude only where it earns it.
How baro generated 808 NestJS Jest tests autonomously in 71 minutes
One prompt: a full Jest unit suite for every service, controller, guard, listener, scheduler, mapper, decorator, filter, helper, and validator under src/. 33 stories, 808 tests, 83.5% branch coverage. Halfway through it hit Anthropic's usage cap. It paused, resumed in the morning, finished — and the L6 quality gate filed exactly zero phantom bug issues.
Benchmarking Claude (via Claude Code) and GPT-5.5 (via Mozaik) in the same parallel-agent run
baro can now drive either Claude or GPT under the same orchestration. Side-by-side run on the same goal, same starting branch, same DAG depth. GPT-5.5 (through Mozaik) finished in 2:16; Claude Opus 4.7 (through Claude Code) in 4:01. The PRs and the per-call latency tell the rest of the story.
I tested Claude Code's new /goal feature against my parallel agent setup. Round one wasn't even close.
Anthropic shipped /goal in Claude Code yesterday — type a goal, walk away, come back to a PR. baro does almost the same thing differently, on top of Mozaik. I ran them side-by-side on a real refactor. /goal won the first round. Then I added one upstream agent — 200 lines, because of Mozaik — and the rematch wasn't close.
Getting the Maximum Out of My Claude Code Subscription
How baro turns one Claude Code subscription into a parallel, autonomous agent runtime. Built on Mozaik. ~6,600 npm installs.