AI Agents Accelerate Development, Shifting Bottlenecks to Orchestration
TL;DR
- Multi-agent engineering accelerates software development by 2-10x, shifting bottlenecks from execution to code review, QA, and async coordination, necessitating new orchestration tools.
- "Context as code" involves curating reusable "context nuggets" (markdown files) for AI agents, enabling just-in-time retrieval for specific tasks rather than overwhelming context windows.
- Agentic workflows benefit from "context engineering" where agents use command-line interfaces (CLIs) to retrieve data, schemas, or run SQL queries dynamically, avoiding duplication in context files.
- Agor, an agent orchestration platform, provides a spatial, multiplayer workspace that manages Git worktrees and live dev environments, enabling collaborative AI-driven development.
- Shared development environments, managed by Agor, allow multiple users and AI agents to collaborate on live codebases, transforming code review and collaborative debugging processes.
- Agent orchestration platforms like Agor can act as internal MCP servers, allowing agents to schedule, monitor, and coordinate other agents, enabling complex automated workflows.
- Security and isolation are critical in multiplayer agent environments, requiring careful management of user permissions and sandboxing to prevent unauthorized access to secrets or system manipulation.
Deep Dive
The integration of AI development agents is fundamentally altering software engineering velocity, shifting bottlenecks from execution to orchestration and collaboration. This evolution necessitates new tooling that supports multi-agent, multiplayer workflows, enabling teams to manage complex AI-driven development processes and unlock significantly accelerated delivery cycles.
The core argument is that AI agents, capable of handling a vast majority of software engineering tasks from coding to research, have dramatically increased individual and team execution speed. This acceleration, however, exposes new bottlenecks. Previously, the bottleneck was the human capacity to write code; now, it is the human capacity to orchestrate, review, and coordinate multiple AI agents. This shift is evident in areas like code review and quality assurance, which, while also being accelerated by AI, become critical points of friction when execution velocity increases by orders of magnitude. The need to manage numerous AI sessions, maintain context across them, and ensure seamless collaboration among both humans and agents has led to the development of specialized orchestration platforms.
The implications of this shift are profound. First, the boundary between data engineering and AI engineering is blurring, with "context engineering" becoming a crucial new discipline. Effectively providing AI agents with the right context, whether through structured markdown files or just-in-time retrieval via CLIs, is paramount. Second, the concept of "context as code" is emerging, where contextual information is treated as a manageable asset, similar to code itself. This enables teams to systematically manage and reuse context, crucial for AI agents that operate with finite context windows. Third, the development environment itself is transforming. Tools like Agor, an agent orchestration platform, are emerging to manage complex multi-agent, multiplayer workflows. These platforms provide visual interfaces, manage git worktrees, templatize prompts by workflow zones, and support session forking and sub-sessions, effectively creating shared, live development environments. This multiplayer aspect is critical for team collaboration, allowing for shared AI sessions, live code reviews within these environments, and seamless handoffs between human and AI collaborators. Finally, the security and management of these shared environments become paramount. As agents gain access to tools and data, robust sandboxing, access controls, and clear role-based access management are essential to prevent unintended consequences and maintain system integrity. The potential for agent-to-agent communication and automated workflows within these platforms suggests a future where complex development tasks are managed through intricate, AI-driven orchestration.
The key takeaway is that the rapid advancement of AI agents is not merely an augmentation of existing workflows but a fundamental restructuring of software development. Teams that embrace this shift, adopting new orchestration tools and adapting their collaborative practices, will gain a significant velocity advantage. Those that do not will face increasing friction as their execution speed outpaces their ability to manage and coordinate these powerful new capabilities.
Action Items
- Create context nuggets: Define reusable Markdown files (<500 lines) for specific tasks (e.g., testing guidelines, front-end conventions) to improve AI agent context retrieval.
- Audit agent execution environments: For 3-5 shared development environments, implement Unix-level isolation to prevent cross-user access to SSH keys and API keys.
- Design agent orchestration policies: Establish limits (e.g., max 12 agents per board) to prevent runaway agent chain reactions and manage context explosion.
- Implement shared development environments: Provision a beefy EC2 instance to host Agor, enabling 12+ engineers to collaborate on live Git worktrees and AI sessions.
- Track agent session metadata: Capture token costs, Git transitions, and prompt history for 3-5 agents to analyze prompting styles and identify efficiency patterns.
Key Quotes
"Generative ai models have grown more powerful and are now being applied to a broader range of use cases the lines between data and ai engineering are becoming increasingly blurry the responsibilities of data teams are being extended into the realm of context engineering as well as designing and supporting new infrastructure elements that serve the needs of agentic applications"
Max Beauchemin explains that the increasing capabilities of generative AI are blurring the lines between traditional data engineering and AI engineering. This shift requires data teams to expand their responsibilities into "context engineering" and support new infrastructure for agentic applications.
"Orchestration is the foundation that determines whether your data team ships or struggles ETL, ML model training, AI Engineering, Streaming - Prefect runs it all from ingestion to activation on one platform."
This quote highlights the critical role of orchestration in data teams' success. Prefect is presented as a platform capable of handling diverse workloads, from ETL to AI engineering, ensuring that data operations can be managed effectively from start to finish.
"Teams that use Bruin need less engineering effort to process data and benefit from a fully integrated data platform."
Bruin is described as an open-source framework that simplifies data integration. The author argues that by handling data movement, lineage, quality, and governance, Bruin reduces the engineering effort required for data processing and provides a cohesive data platform.
"The bottlenecks have shifted very very quickly right so if you you know people report being multiplied by ai by different degrees some people will say like i'm 20 faster i'm slower personally like since the get go i've been like two five 10x like somewhere i would say between two x and 10x on coding specifically or you know the act of you know designing and shipping software some of the areas that are still you know that are kind of becoming a traffic jam now are you know code review for sure qa right so you still need to go in your app though you can get a lot of help from the agents to write good unit tests and integration tests"
Max Beauchemin observes that as AI accelerates software development, traditional bottlenecks like code review and QA are becoming more prominent. While AI can assist in writing tests, the human elements of review and quality assurance still present challenges in the accelerated workflow.
"Treating context as code is like you know the first thing to do probably and so that's definitely an important thing to do um to do early on and systematically now"
This statement emphasizes the importance of managing contextual information for AI agents as if it were code. Beauchemin suggests that organizing and treating this context systematically is a foundational step for effective AI-driven development.
"Agor is a tool where you can essentially manage the tool will manage your git work trees for you maybe we'll start from the ground up so first thing is you add your repos to agor and it's a super visual interface right so you you get into this agor interface and the first thing you would do is bring in the work the the the git repos you want to work with and then it will manage your work trees for you"
Max Beauchemin introduces Agor as an agent orchestration platform that visually manages Git worktrees. The core functionality involves users adding their Git repositories to Agor, which then handles the creation and management of worktrees for feature development.
"The beauty in agor is that you can really just define your own framework it's an open canvas you know you can define whatever zone you want and whatever prompts you want and whatever whatever workflow you want"
Beauchemin highlights Agor's flexibility, describing it as an open canvas where users can define their own frameworks, zones, prompts, and workflows. This allows for highly customizable agentic engineering processes tailored to individual or team needs.
"The meta stuff was kind of a discovery along the way right so early on i was like hey it'd be cool if agor exposed this mpc serve like this mpc agor service to itself so it can be used internally you know but externally too so since i've had that i added a lot of mpc tools right so get the list of session get the list of work trees and inspect the session you know fire up an environment look at the logs for an environment you know create a user"
Beauchemin discusses the development of Agor's internal MCP (Multi-Command Protocol) service as a key feature. This allows agents within Agor to interact with the platform itself, enabling functionalities like listing sessions, inspecting environments, and creating users, which was an emergent discovery during development.
Resources
External Resources
Books
- "The Hug" by The Freak Fandango Orchestra - Mentioned as the source of intro and outro music for the podcast, with a Creative Commons license.
Articles & Papers
- "Love, death and a drunken monkey" (freemusicarchive.org) - Mentioned as the source of intro and outro music for the podcast.
Tools & Software
- Agor - An open-source agent orchestration platform for managing multiple AI agents and collaborative development environments.
- Apache Airflow - An orchestration tool for data pipelines.
- Apache Superset - An open-source data visualization and exploration platform.
- Bruin - An open-source framework for data integration, movement, lineage tracking, data quality monitoring, and governance enforcement.
- Claude Code - An AI coding assistant.
- Codex - An AI coding assistant.
- Datafold's Migration Agent - An AI-powered solution for data migrations.
- GitHub Codespaces - A hosted development environment service by GitHub.
- Git Worktrees - A Git feature that allows multiple working trees to be associated with a single repository.
- Helm - A package manager for Kubernetes.
- Opencode.ai - An open-source AI coding platform supporting multiple models.
- Playwright MCP - A browser automation tool for AI agents.
- Prefect - An orchestration platform for data workflows, including ETL, ML model training, and AI engineering.
- Tmux - A terminal multiplexer for managing multiple terminal sessions.
People
- Maxime Beauchemin - Guest, interviewed about multiplayer, multi-agent engineering and its impact on data and AI systems.
Organizations & Institutions
- Anthropic - Provider of the "Anthropic Pro Max plan" used for AI coding agents.
- Cash App - Relies on Prefect for data orchestration.
- Cisco - Relies on Prefect for data orchestration.
- Data Engineering Podcast - The podcast hosting the interview.
- Dbt Cloud - Offers a credit for migrating to Bruin Cloud.
- GitHub - Provider of GitHub Codespaces and the Git version control system.
- Hugging Face - Investing in agentic environments.
- Ona - Associated with Gitpod, focusing on agentic coding environments.
- Preset - A company where Maxime Beauchemin works and has adopted AI coding agents.
- The Freak Fandango Orchestra - Creator of the podcast's intro and outro music.
- Whoop - Trusts Prefect for their data operations.
- 1Password - Trusts Prefect for their data operations.
Websites & Online Resources
- agor.live - The website for the Agor agent orchestration platform.
- dataengineeringpodcast.com/bruin - Link for information on Bruin.
- dataengineeringpodcast.com/datafold - Link for information on Datafold.
- dataengineeringpodcast.com/prefect - Link for information on Prefect.
- freemusicarchive.org - A repository for free music.
- linkedin.com/in/maximebeauchemin - Maxime Beauchemin's LinkedIn profile.
Other Resources
- AI Engineering - A field transforming how individuals and teams build data and AI systems.
- Context as Code - The concept of treating context information as code for AI agents.
- Context Nuggets - Markdown files containing reusable context information for AI tasks.
- Data Migrations - A process that can be accelerated by AI tools.
- ETL (Extract, Transform, Load) - A traditional data processing method.
- Git Branches - Traditionally private, but evolving with collaborative development environments.
- Git Worktrees - A Git feature enabling multiple working directories for a single repository.
- Just-in-Time Retrieval - A method for agents to gather necessary information without bloating context windows.
- MCP (Multi-modal Command Protocol) - A system allowing agents to schedule, monitor, and coordinate other agents.
- Multi-Agent Engineering - The practice of using multiple AI agents to build systems.
- Multiplayer Engineering - A collaborative approach to building systems with AI agents.
- Multi-modal Context - AI's ability to process and understand information from various sources, including visual data.
- Spatial First Environment - A design approach for interfaces, like Agor, that uses a 2D canvas for organizing projects.
- State Persistence - The ability of systems to retain data and configuration across sessions.
- Templated Prompts - Pre-defined prompts used to guide AI agents for specific tasks or zones.
- Unix-level Isolation - A security measure to separate user environments.