How it works

BellowAI is a multi-agent pipeline running on cloud infrastructure. You never manage any of it — but here is what happens between your prompt and your plan.

The path a run takes

When you run a session, the work flows through a handful of stages designed for speed, isolation, and cost control:

1. Your browser sends the prompt and your model choice to the BellowAI API.
2. The API validates the request and places the job on a message queue rather than processing it inline — so the request returns quickly and the heavy work happens out of band.
3. An ephemeral worker picks the job up off the queue. It is spun up to do exactly one job and torn down the moment it finishes — nothing lingers between runs.
4. The worker calls the LLM (your chosen Anthropic or OpenAI model, using your key) and drives the agents through their stages.
5. Results stream back as they are produced. Output is written incrementally and streamed to your browser, so you see the plan build in real time instead of waiting for one final response.

Because the worker is ephemeral and per-job, compute spins up only when there is work and disappears when the job is done — which keeps the system cost-efficient and keeps each run isolated from the next.

The agents in the pipeline

The pipeline is a sequence of specialized agents. Each one has a focused job and hands clean, structured output to the next:

• Planning — analyze the idea, clarify scope, and frame the problem.
• Epics & features — decompose the plan into epics and the features that deliver them.
• Stories — turn features into implementation-ready user stories with acceptance criteria.
• Development & validation — draft the technical approach and validate the plan for gaps before you ship it.

You review and approve at each gate, so a human stays in control of what moves forward.

Parallel Agent Dispatch

For larger plans, BellowAI can run independent work at the same time instead of strictly one stage after another. This is Parallel Agent Dispatch (PAD), and at a high level it has three roles:

• A planner looks at the work and decomposes it into slices that can be worked on independently.
• Each slice is executed on its own, producing its portion of the plan and streaming results as it goes.
• A coordinator fans the slices out, runs a conflict check as they come back, and merges the results into one coherent plan.

The payoff is shorter wall-clock time on big plans without losing coherence — the conflict check is what keeps parallel work from contradicting itself when it merges back together.