A multi-agent build pipeline
“I want several agents planning, coding, reviewing and shipping in parallel instead of one agent doing everything serially”
Fleets fail on coordination, not capability. The stories that work all have the same two things: a single screen where you can see what every agent is doing, and a gate the work has to pass before it lands. So the orchestrator and the review surface come first, together — an orchestrator without review just produces bad work faster. Checkpointing matters more than it sounds like it should, because long parallel runs fail partway and restarting from zero is what kills the economics. Plan enforcement is the last piece and the one people add after their first agent confidently does the wrong task well.
Real posts from the Hermes community, mirrored from Nous Research's user stories corpus.
Day 8 of Building AGI for my Hermes Agent: Auto-Build saved me loads of time and tokens. Main agent (GPT-5.4) breaks a plan into phases, coder agent (MiniMax M2.7) implements, QA agent (local Qwen 35B A3B) tests. Plan → implement → test → fail → repair → ship.
18 agents working in parallel, 97 completed tasks. Scout agents hunt real pain points, an orchestrator scores against a rubric — anything under 65/100 gets shelved — then a builder+tester ship a working CLI tool. The system self-healed a broken deliverable on…
WHAT IS THE NEW KANBAN FEATURE BUILT INTO HERMES? (IT'S GAME CHANGING) — image post showing Hermes' new built-in Kanban board where a parent agent posts cards and child subagents pull them, work in parallel, and report back, turning the agent into a…
Made a network based kanban for agents with local client to minimize token use. Comes with web ui so meat based agents can see what is going on.
Run three fully isolated agents — a programmer (uses Claude Code), a deep researcher, and a designer — each with its own personality, memory, skills, and Telegram bot. Hermes packages a gateway around a learning agent.
- Parallel agents multiply spend linearly and mistakes faster than that. Pair this stack with the cost-visibility one before scaling past a handful of lanes.
- Several of these tools assume they own the session lifecycle. Running two orchestrators at once is a common and confusing failure.