Race to the Best: Run Parallel Agents with agenthub

Where this fits

You have a slow parse() function in a critical service, and three plausible optimization strategies. If you try them one by one, you burn time switching contexts and carrying over mental state. agenthub starts from that scenario: define the task once, spawn N subagents in isolated git worktrees, let each pursue a different strategy in parallel, then evaluate results and merge the winner.

Why agents benefit

• Each subagent gets a full git worktree, so it can edit code, run tests, and commit independently without colliding with its peers.
• The coordinator exposes hub-status with DAG state, agent progress, and branch status, giving an agent an observable way to monitor the race and decide if a spawn needs more time.
• An LLM judge or metric can rank branches on criteria you set, which removes the "last agent to finish wins" bias and makes evaluation repeatable.
• Multiple approaches are explored in parallel, increasing the chance that a non-obvious solution—like a cache strategy vs. a query rewrite—actually gets tried.
• Losing branches are archived automatically after merge, keeping the repo history clean while preserving the work.

Practical scenarios

Code optimization race

Spawn four agents on a latency problem: caching, SQL query rewrite, algorithm change, dependency removal. Run the test suite and benchmark in eval, merge only the branch that passes and has the lowest p95 latency.

Content variation test

Ask for three landing page variants aimed at the same persona. Each agent writes in its own branch with a different hook; an LLM judge scores on tone, clarity, and call-to-action strength. The winning copy is merged into main.

Research exploration

Task a small fleet with "compare Postgres vs. ClickHouse for our analytics workload." Each worktree contains a scratch/ file where agents collect metrics and tradeoffs in isolation. The judge ranks on completeness and factual grounding; the best report is merged for the team to review.

Add it to your agent workflow

Install with:

npx skills add alirezarezvani/claude-skills --skill agenthub

Then inside any git repository, start a session:

/hub:hub-init task=\"optimize parse() in src/parser.ts\" agents=4 criteria=\"all tests pass and lowest p95 latency\"
/hub:spawn
/hub:eval
/hub:merge

/hub:hub-init records the task and eval criteria, /hub:spawn creates isolated worktrees, and /hub:merge brings the winner back into your current branch. Requires a git repo; nothing about the rest of your stack changes.

Compare before adopting

Before making this a default step in every task, compare quality signals: does the evaluator actually align with your definition of "best"? Maintenance freshness: the skill is iterating (v2.1.2, updated 2026-03-17), but you need confidence it stays compatible with your agent runtime. Alternatives: running the same work manually through separate branches, or using a general orchestration framework, may be enough for simpler tasks. Workflow fit: if your task needs tight collaboration between agents rather than competition, a parallel race is the wrong shape.

Why it is worth tracking

With 25k stars and a focused model for divergence-and-merge, agenthub is a strong signal that parallel agent competition is moving from experiment to practical workflow. It is worth tracking because it makes a previously manual pattern—multiple branches, multiple attempts, judging the result—explicit and observable. Evaluate it when you have a task that rewards trying many approaches in parallel and you already trust git as the substrate. For a maintained view of skills in this ecosystem, see the OpenAgentSkill entry at https://www.openagentskill.com/skills/alirezarezvani-agenthub.