Autonomous Coding Agents in Software Delivery Platforms
This article discusses Harness's acquisition of Augment Code, integrating autonomous coding agents into their software delivery platform. It highlights how these agents aim to automate code generation, testing, and even issue remediation by leveraging deployment history and feedback from downstream systems. The core system design challenge lies in orchestrating these AI agents, managing context across various stages of the SDLC, and ensuring secure, controlled automation.
Read original on The New StackHarness's acquisition of Augment Code introduces autonomous coding agents that aim to revolutionize the Software Development Lifecycle (SDLC). The goal is to move beyond simple code generation to creating valuable software that reaches customers by integrating AI into testing, deployment, and even remediation processes. This represents an evolution in CI/CD pipelines towards more intelligent, self-correcting systems.
The Cosmos Software Factory Agent
The Cosmos agent is designed to take a requirement or bug report and produce a pull request, handling code writing and testing within isolated virtual machines. A key system design aspect is its ability to learn from feedback: it can pick up failed checks and review comments, using this information to make further changes. Teams can customize these agents for specific workflows, indicating a need for flexible configuration and extensibility in the agent's architecture.
Architectural Consideration: Agent Orchestration
The integration of AI agents into the SDLC requires robust orchestration mechanisms to manage their lifecycle, allocate resources (like isolated VMs), and coordinate their actions across different development stages, from coding to testing and deployment.
Contextual Understanding and Feedback Loops
A critical, yet currently unshipped, feature is the integration of Augment's Code Context Engine with Harness's Software Delivery Knowledge Graph. This aims to give Cosmos agents access to rich historical data, including deployment failures and security findings. This closed-loop feedback system allows agents to automatically investigate and fix issues by leveraging past failure contexts, reducing developer intervention. Designing this integration requires careful consideration of data schemas, retrieval mechanisms, and the secure exchange of sensitive information between potentially disparate systems.
Proposed Feedback Mechanism
Downstream Agent (e.g., Testing, Security) -> Identifies Issue + Context -> Sends to Cosmos Agent
Cosmos Agent -> Accesses Software Delivery Knowledge Graph (Deployment History, Security Findings)
Cosmos Agent -> Generates Fix (Code Change) -> Creates Pull Request
Developer -> Reviews & Approves/Rejects PR
(If Rejected/Failed) -> Feedback Loop Continues to Cosmos Agent with new contextSecurity and access control are paramount. The article raises questions about how agent permissions will work across the integrated Harness platform, especially concerning access to production data and the actions agents can take. A robust identity and policy management system is essential to prevent unauthorized access or actions by autonomous agents, particularly given past incidents where coding agents inadvertently exposed sensitive data.