Introduction (Who/What) In this scholarly technical discourse, host David Andre interviews Mickey, a senior software engineer at Convex, to dissect the operational paradigms of "Agentic Engineering" in 2026. Mickey, who relies on autonomous artificial intelligence to generate approximately 95% of his production grade code, outlines the critical shift from casual "vibe coding" to highly structured, deterministic, and automated multi agent workflows. The conversation centers on building sustainable, highly scalable code architectures, optimizing model context inputs, and accelerating software deployment cycles. 🤖
Structured Summary
-
The AI Tech Stack 🛠️ The contemporary developer's pipeline hinges on a clean separation between the predictive linguistic core (the model) and the operational wrapper (the harness). Mickey's primary development harness is Cursor, which coordinates specialized tools, APIs, system instructions, and customized markdown configurations. This wrapper converts human intent into tokenized coordinates to guide autonomous agents. Cursor's proprietary harness out-performs alternative environments on developer benchmarks by streamlining automated tool calls. He utilizes a dual model approach: OpenAI's GPT-5.5 (extra high fast) serves as the primary engine for complex architectural logic and back end integration, while Anthropic's Opus 4.7 Max is leveraged exclusively for front end user interface modifications. This strategic allocation maximizes model specific strengths while completely bypassing performance lag.
-
Context Engineering & Code Structure 📂 Effective agentic engineering rejects "vibe coding" in favor of rigorous context minimization. Since excessive context inflation degrades model reasoning, keeping the context window narrow and highly precise is vital. Mickey minimizes his reliance on heavy
agent.mdsystem prompts, trusting modern models to autonomously detect tech stacks directly from raw files. Developers utilize Vercel’s open source library (npx open-source) to fetch the literal source code of external dependencies and dump them directly into a designated repository folder. This provides the agent with the absolute source of truth rather than obsolete, human written documentation. Furthermore, Mickey emphasizes utilizing a disciplined "service layer" to decouple repeated runtime mechanics from core domain logic. After generating any new feature, running specialized architectural skills ensures duplicate codebase elements are refactored into modular, reusable functions, preventing technical debt, long-term structural degradation, and "code smell." Additionally, choosing back end technologies like Convex—which are written entirely as TypeScript code rather than managed via dashboards—ensures the agent maintains complete visibility of the system schema. -
The Grep Loop Workflow 🔄 The core of Mickey’s rapid development cycle is the automated feedback loop, exemplified by Greptile’s "Grep Loop" system. Once a Pull Request (PR) is initiated, Greptile assesses code quality, security vulnerabilities, and logic flaws, assigning a rigorous confidence score from one to five. If the PR receives an imperfect score, the developer executes the
/grep-loopcommand. This prompts the background agent to ingest the feedback, resolve the specified issues, push automated fixes to GitHub, and trigger a fresh review cycle automatically. Advanced reasoning models write localized tests to validate their fixes, executing repeatedly in parallel threads until a perfect score of five is achieved. This rigorous, self healing QA process runs asynchronously, freeing human developers to focus on higher level system design. -
Mindset, Security, & Future Outlook 🌐 Success in the agentic era requires a psychological pivot toward radical bias for action and highly proactive deployment. Drawing inspiration from Silicon Valley's "delusional" optimism, developers are urged to launch imperfect Minimum Viable Products (MVPs) immediately to capture market feedback, rather than over-engineering minor features in isolation. However, this velocity demands uncompromising cybersecurity protocols. In an era of automated exploit agents, Mickey advocates for strict package verification—such as instructing agents to never install dependencies under 14 days old to mitigate zero day supply chain attacks. Standard practices must include hardware based Two-Factor Authentication (2FA), encrypted password managers, and vocal passphrases to combat highly realistic, automated AI voice cloning schemes. Ultimately, the next quarter will see a massive democratization of knowledge work, where non technical operators leveraging these precise integration frameworks can easily outbuild traditional software engineers.
Academic Takeaway 🎓 Agentic engineering redefines the developer from a manual syntax writer into a strategic system architect. By utilizing a highly optimized model harness, maintaining rigorous context boundaries, and enforcing automated validation loops, engineers can construct resilient, enterprise grade software at unprecedented speeds. The future of technology belongs not to those who merely write code, but to those who master the orchestration of the autonomous AI agents writing it.