.svg)
When an engineering manager at Augment Code ran their first eval in week one and shipped a new model in week two, it wasn't a fluke.
It was proof that AI coding tools have finally cracked the context problem that's plagued large, messy codebases - the exact places where 80% of enterprise value gets created.
In the latest AI Giants podcast episode, Codacy CEO Jaime Jorge spoke with Vinay Perneti, VP of Engineering at Augment Code, about building AI tools that actually work in the real world of legacy code, technical debt, and repositories with millions of lines. Their conversation revealed why most AI coding tools fail at scale and how Augment's radical approach to context management changes everything.
What is Augment Code?
Augment Code builds AI coding assistants specifically designed for large, complex codebases. Unlike tools that rely on simple text search, Augment ingests entire repositories, creates semantic embeddings, and maintains millisecond-level sync with code changes. They're now launching their context engine as a standalone API.
TLDR; What Engineering Leaders Need to Know
- Context is everything - Augment ingests entire codebases and stays synced in milliseconds, not through grep and markdown files
- Onboarding time cut from 4-5 months to 6 weeks - New engineers ship complex PRs touching entire codebases in their first month
- The "contractor vs. employee" model - Augment provides context (like an FTE), while using best-in-class models for intelligence
- 60-80% code review acceptance rate - Their AI suggestions are accepted more often than many human reviews
-
Roadmaps are dead - Augment threw out their 2025 roadmap on January 25th; now they plan in quarters
- Automation bias is cultural, not technical - The person pushing the PR owns the code, period
The Death of Traditional Onboarding
When engineers join companies with large codebases, they spend weeks trying to balance asking "stupid questions" versus appearing incompetent. They open hundreds of PDFs with meaningless names just to build a timeline. They take 4-5 months to become productive.
Augment destroyed this model.
"At Augment in week six, they've already put out a pretty complex code PR that touches like a wide range of the codebase"
One engineer told him the tool let them "ask all the stupid questions" without bothering teammates, getting them 80% of the way to answers before human conversations even started.
The engineering manager example is even more striking - joining with no context, running evals in week one, shipping models in week two. "That is unheard of," Perneti emphasizes.
Why Context Beats Intelligence
Augment's philosophy challenges Silicon Valley's obsession with model intelligence. They view it through an unexpectedly practical lens: "Contractors are just borrowing intelligence, but they're missing context. Full-time employees typically have both."
This insight drove their three-pronged context strategy:
- Research-driven embeddings - Custom embedding and retrieval models trained in pairs for maximum quality
- Expanding context sources - Recently added PR history because "adding a feature flag touches like 20 different places"
- Context as an API - Soon allowing companies to add their own context sources programmatically
The result: While competitors offer 20+ models for users to choose from, Augment provides just three carefully selected options.
"Let us solve the hard problems of what models make sense," Perneti explains. "You shouldn't be spending mental cycles on how do I get the best context."
The Pricing Revolution Nobody Wanted
Perneti also addressed their recent pricing overhaul that "upset some people." The original model charged per user message regardless of complexity - asking "what does this method do" cost the same as complex multi-file refactoring.
"The agents started doing more and more work for one user message, and the pricing just didn't make sense," he admits. Their prompt enhancer feature, which automatically enriches simple prompts with codebase context, made the problem worse by making each message do exponentially more work.
Their solution: transparent, usage-based pricing aligned with actual compute costs. "Whenever you need to make a hard decision, the best thing to do is be transparent," Perneti notes.
Code Review's Unexpected Evolution
Augment's code review tool achieves 60-80% acceptance rates - often higher than human reviewers. But Perneti sees it differently than replacing humans: "What are some things that AI can be really good at and let AI do that? And what are some things that as a human we would have to be doing?"
The future isn't AI checking AI in infinite loops. It's about intent review and shared understanding as codebases evolve rapidly. The AI catches the technical details that would require "a lot of thinking cycles," freeing humans to focus on direction and architecture.
The Bottom Line
As software development transforms from implementation to orchestration, the companies that win won't be those with the smartest models - they'll be those who solve the context problem. Because in the real world, developers don't work on pristine greenfield projects. They work on large, messy, critical codebases where one mistake can cost millions.
Augment's approach - treating AI as a tool, not a replacement, while solving the fundamental context problem - points to where the industry is heading: not fewer developers, but developers working at a fundamentally different level of abstraction.
This is why Codacy built Guardrails - to ensure AI-generated code in these complex environments maintains security and quality standards. When AI agents are touching 20 different parts of your codebase for a single feature flag, systematic safeguards aren't optional.
