Essential AI Agent Workflows: Context Management and Prompting Best Practices

Episode 1 of "Build Crew Live" video features the Amp team introducing their new community, Build Crew, and discussing their personal best practices for prompting and working with AI agents.

Launching the Build Crew Community

The core of the episode is the announcement of Build Crew, a free community for developers shipping with AI agents, launched by the Amp team.

• Purpose: The community aims to connect developers, share knowledge, and provide support for the challenges of building with agents.
• Features: Joining gives access to a private Twitter DM group, a leaderboard, and badges, and offers a $100 credit for those who use Amp.
• Shared Learning: The team hopes sharing public threads will become a standard way to move the industry forward and combat "imposter syndrome" by normalizing the learning process.

AI Agent Best Practices & Workflow Strategies

The Amp team shared several practical and advanced techniques for effective agent interaction:

• Provide a Feedback Loop: The agent can do a much better job if it can reproduce a bug or get feedback on its code, similar to how a developer uses simple scripts to reliably reproduce an issue.
• The "Desk" Analogy for Context: Think of the context window as your desk. Keep all necessary information there, but move high-level, complex questions over to the "senior engineer"—the Oracle (a smart reasoning sub-agent)—to prevent cluttering your main thread's context.
• Handling Long-Running Processes: Use a terminal multiplexer like tmux to start processes that wait for user input (like a CLI), allowing the agent to remote control the terminal and interact with the process.

Structuring Prompts for Success

Effective prompt structure is key to managing complexity and ensuring clarity:

• Goal, Phases, and Confirmation: Start with a clear goal (TLDR), break the task into phases, and ask the agent to repeat its mission back to you before coding. This acts as a proofreading/chain-of-thought mechanism to ensure alignment.
• External Memory for Context: When starting a new thread or reaching a context limit, ask the agent to summarize all the relevant information (problem, background, current state, next steps) and write it into a markdown file. A new thread can then simply reference this file, maintaining long-term memory.
• Forking Threads: Use the "fork" feature to freeze the context of a thread and create variations from a specific message point. This lets you experiment with different approaches without modifying the core context.

The Team's Pro-Tips for Developers

The team concluded with quick-fire advice for improving agent development:

• Slow Down and Plan: Take the time to think through the problem and let the agent plan the solution. This slower approach often leads to smoother, faster results.
• Confirm the Fix: When asking the agent to fix a bug, instruct it to confirm the fix with a new test or a reliable confirmation step.
• "Don't Get Drunk on Tokens": Even with massive context windows, don't try to fill them up. Keep your threads small, fork, and compact when necessary to reduce noise and maintain focus.
• Ask for Anything: Do not limit what you ask an agent to do; you would be surprised at what the latest models are capable of achieving.

Full Video: Watch on YouTube(opens in a new tab)

Build Crew: https://buildcrew.team(opens in a new tab)