How to Build an Agentic Engineering Workflow

TL;DR

Build an agentic engineering workflow around a unified interface, visible agent states, priority-based task handling, and subscriptions that maximize token access. Use local agents for convenient development, but consider cloud agents when workloads require greater scale, isolated environments, persistent sessions, and resilience when a laptop closes or briefly loses connectivity.

Transcript

My name is David Andre and over the past three years, I've spent well over 2,000 hours coding with AI. And I've also personally interviewed some of the most legit and productive people in the space of Agentic Engineering. And in this video, I'm going to show you my full Aentic engineering setup as it currently stands, holding nothing back. Starting... Read More

Key Insights

• BB is an open-source, free interface that places multiple agents, models, subscriptions, projects, and subagents inside one user interface. Its main advantage is flexibility, because users are not limited to the models or subscription offered by a single proprietary coding application.
• CMAX is useful for splitting a workspace into terminal panes, launching different agent harnesses, and using a built-in browser. Its left-sidebar design becomes difficult to manage when projects contain many agents and workspaces, making it better suited to smaller workloads with only a few concurrent activities.
• Herder is a lightweight terminal-based runtime for starting agents and sessions while tracking their states. Its interface marks agents as done, idle, blocked, or running, allowing users to monitor parallel work without repeatedly opening every session or watching each agent continuously.
• Human attention is the limiting resource in multi-agent engineering workflows. As the number of workers grows, the user should interact with a manager agent and rely on clear state tracking instead of maintaining separate conversations with every individual worker agent.
• Coral is a priority-based interface designed to order completed agent work by importance. A finished P1 agent moves ahead of a P4 agent, preventing the user from responding to low-priority work while a more consequential result is waiting for review.
• Subscription value depends on maximizing tokens while minimizing cost. The recommended starting combination for a $30 monthly budget is OpenCode Go plus ChatGPT Codex, while larger budgets can add Claude Code, Cursor, or higher-tier plans according to available funds.
• API pricing is described as the least economical option because it lacks the subsidy and increased token allowance available through subscriptions. The workflow therefore combines several subscriptions based on budget instead of paying directly for API usage whenever subscription access is available.
• Cloud agents provide scalability, isolated environments, persistent sessions, and durable access to internet and electricity. They address limits of local machines, but hosted services can create ecosystem lock-in after users configure environments, store secrets, accumulate sessions, and commit context to one provider.

Questions & Answers

Q: How do you build an agentic engineering workflow?

Build the workflow around four elements: a unified interface, a lightweight agent runtime, visible state tracking, and priority-based review. BB can consolidate different agents, models, subscriptions, projects, and subagents. Herder can show which agents are running, idle, blocked, or finished. A manager agent should coordinate workers, while a system such as Coral can surface higher-priority results first.

Q: What is BB used for in agentic engineering?

BB is used as a single interface for working with different subscriptions, models, agents, and projects. It supports Codex, Claude Code, Pi, Cursor CLI, OpenCode, Grok, and Hermes within the same user interface. It can also launch subagents through a parent manager agent, and users can inspect those subagents without manually writing or managing every prompt.

Q: Why is agent state tracking important?

Agent state tracking shows whether each worker is done, idle, blocked, or running. This lets the user direct attention toward sessions that require action instead of constantly checking every agent. The workflow treats this visibility as essential because human attention is limited, especially when a manager agent is coordinating numerous worker agents across several projects and concurrent tasks.

Q: When should you use CMAX for coding agents?

CMAX is suitable when you need a terminal workspace divided into several panes, want to launch different agent harnesses, or benefit from a built-in browser. It works well when only a few agents and activities are running. The transcript advises against relying on it for larger projects because its left sidebar becomes difficult to manage across many agents and workspaces.

Q: How does Coral prioritize completed agent work?

Coral assigns different importance levels to agents and orders finished work accordingly. When a P1 agent completes its task, its result should rise above work from a P4 agent. This prevents the user from answering or reviewing a low-priority session while a more important result is waiting, turning agent completion into an ordered queue rather than a random sequence.

Q: Which AI coding subscriptions provide the best value?

The recommended selection depends on budget. For $30 per month, the transcript recommends OpenCode Go and ChatGPT Codex. At $50, it adds a $20 Claude Code subscription, and at $70 it adds a $20 Cursor subscription. At $110, it recommends choosing one larger plan, with ChatGPT Codex described as offering a better deal than Claude Code at that point.

Q: Why should agentic engineers avoid API pricing?

API pricing is presented as the worst value because it does not include the subsidy or increased token access associated with subscriptions. The stated objective is to obtain as many tokens as possible for as little money as possible. The workflow therefore recommends combining subscriptions according to budget, rather than defaulting to direct API charges for routine agentic engineering work.

Q: Why are cloud agents useful for software development?

Cloud agents can scale beyond the resource limits of a local computer and provide isolated environments, persistent sessions, and durable access to internet and electricity. Local agents may overload a machine when several workers run database tests or a complete test suite simultaneously. Cloud execution also avoids losing active sessions when a laptop closes or briefly loses its internet connection.

Summary & Key Takeaways

• The workflow uses several complementary interfaces. BB provides one open-source workspace for multiple agents, models, subscriptions, projects, and automatically launched subagents. CMAX offers terminal panes and a built-in browser, while Ghostty and Herder provide a lightweight terminal setup that clearly displays whether each agent is working, idle, blocked, or finished.

• Managing many agents requires more than opening additional terminal panes. Human attention is limited, so a manager agent should coordinate worker agents while the interface exposes their current states. Coral extends this idea by ranking completed work according to priority, ensuring that a finished P1 agent receives attention before a less important P4 agent.

• Subscription selection should maximize available tokens for the budget, with the recommended combination changing as spending increases. The workflow strongly favors subsidized subscriptions over API pricing. Cloud agents add scalability, isolated environments, persistent sessions, and durable access, but proprietary services can create lock-in through environment setup, stored secrets, session history, pricing, and accumulated context.