From Single Bots to Full Agent Orchestration Systems
AI coding agents are software development automation tools that can reason over codebases, coordinate with developer workflows, and execute multi-step engineering tasks across services without manual micromanagement. Most early tools lived inside one IDE or repository, focused on suggestions and snippets. Now, enterprises are pushing towards agent orchestration systems that coordinate many agents, connect to CI/CD, and operate across hundreds of services. This shift is driven by the gap between basic code generation and how large engineering organisations work day to day. They need agents that understand tickets, repositories, and company conventions, and that can open pull requests, track tests, and respond to feedback. Emerging platforms such as GitHub Actions agentic workflows and Block’s internal Builderbot framework show that the next phase of AI coding agents is less about individual assistants and more about connected, policy-aware automation spanning entire software delivery pipelines.
GitHub Actions Agentic Workflows Bring AI into CI/CD
GitHub Agentic Workflows, now in public preview, embed AI coding agents directly inside GitHub Actions so they can act wherever CI/CD already runs. Teams describe automated engineering tasks in natural language Markdown, which GitHub compiles into standard Actions YAML. That means agentic workflows can reuse existing runner groups and policies, inheriting the same security controls that govern other pipelines. According to GitHub, these agents can already handle reasoning-heavy work such as issue triage, pull request reviews, CI failure analysis, and documentation updates. The feature can also rely on the built-in GITHUB_TOKEN, removing a separate token-management step for teams. Early adopters like Carvana and Marks & Spencer are using GitHub Actions agentic workflows to automate work that spans multiple repositories, from vulnerability remediation and dependency maintenance to routine change reviews, and then standardise those automations as reusable workflows across their organisations.
Inside Block’s Builderbot: A Central Brain for AI Coding Agents
Block’s Builderbot framework shows what custom agent orchestration systems look like when tuned to a complex, multi-service environment. Built on the open-source Goose framework, Builderbot acts as a central orchestration layer that coordinates a fleet of AI coding agents across hundreds of services and hundreds of millions of lines of code. Engineers talk to it from a single Slack thread by tagging @builderbot, turning the conversation itself into the development environment. Builderbot handles research, planning, and coding while humans steer in real time. It can pick up tickets from Linear and Jira, create branches, open pull requests, and then watch CI and iterate on feedback, all without leaving chat. Block reports that the system runs more than 200,000 operations a day and merges about 1,500 pull requests a week, accounting for roughly 15% of all production code changes across the company.
Automating Cross-Service Work Across a Massive Codebase
Block originally adopted standard single-repository coding assistants, but these tools could not cope with the mechanical complexity of hundreds of interconnected services. Builderbot was designed to solve that problem by mapping the full structural context of the company’s codebase: it catalogues every active service, documents each internal API, and encodes engineering conventions so agents can modify any repository with the right patterns. Builderbot also has the permissions and context to carry out complex, cross-service tasks, from routine bug fixes to large-scale architectural migrations across multiple databases. A Cash App engineer, for example, can initiate changes in a Square backend service they have never touched, because the orchestration layer supplies the needed context automatically. Block notes that the tool operates solely on source code and system configurations, avoiding customer data and payment information while still enabling large-scale software development automation.
Why Custom Agent Frameworks Are Becoming an Enterprise Pattern
Together, GitHub Actions agentic workflows and Block’s Builderbot signal an emerging pattern: enterprises are building custom AI agent orchestration systems that sit inside their existing tools instead of adding separate AI silos. By integrating with Slack, ticketing platforms, and GitHub, these systems reduce context switching and make AI feel like part of the normal workflow rather than an extra step. Crucially, they also align automation with local architecture and policy. GitHub’s approach lets organisations reuse existing runner groups and security constraints, while Block’s internal framework encodes its own service boundaries, APIs, and conventions. This tight integration gives teams confidence to automate more of the delivery pipeline, from repository maintenance to cross-service feature work. As more organisations reach similar scale and complexity, custom agent frameworks are likely to become a standard way to turn AI coding agents into reliable, end-to-end software development automation.
