Infinite Code Context: AI Coding at Enterprise Scale w/ Blitzy CEO Brian Elliott & CTO Sid Pardeshi

Just-in-time agent generation

Blitzy generates agents dynamically with prompts written by other agents and tools selected contextually, avoiding hard-coded harnesses that depreciate as models improve. This allows the system to automatically adapt to new model capabilities through configuration changes rather than code rewrites.

Cross-model verification

The platform never relies on a single model family, instead using models from different developers to check each other's work. This redundancy catches errors and compensates for individual model limitations in long-running complex systems.

Harness over model philosophy

The team is "bearish on LLM capabilities as pure standalone assets" and "bullish on orchestration," believing AGI-type effects emerge from sophisticated system design rather than waiting for perfect single models.

Schematizing vs. semantic clustering

Instead of relying on semantic similarity, Blitzy creates deep knowledge graphs mapping relational entities (globals, classes, functional relationships) across 100-million-line codebases. This schematization enables precise line-level context injection and extraction.

Effective context window management

Models experience intelligence depreciation after using 20-40% of advertised context windows. Blitzy manages this by strategically loading only relationally relevant context for specific tasks while pulling out unnecessary information to maintain quality.

Language-agnostic understanding

The platform builds an AST-like abstraction layer that understands code relationships agnostic of programming language, enabling analysis across polyglot enterprise environments without version-specific parsers.

Parallel production environments

Blitzy spins up actual running instances of client applications in isolated cloud environments to observe runtime behavior, compile-time dependencies, and production logs—going far beyond static analysis to understand how code actually executes under load.

Recursive build validation

The system handles legacy enterprise apps with complex manual build processes, database dependencies, and credential requirements, iteratively discovering missing packages and dependencies to create correct build instructions even when documentation is absent.

Screenshot-based QA

Agents validate code by actually running applications and capturing screenshots of UI interactions, creating a recursive correction loop that verifies not just compilation but functional behavior in production-like conditions.

Taste-based outcome evaluation

Rather than testing isolated functions, Blitzy evaluates against complete real-world applications (like Apache Spark) measuring progress toward 100% human-level completion. This requires "extreme taste" to judge whether 90% vs 95% completion meets enterprise standards.

Memory over fine-tuning

The team is more bullish on advances in AI memory systems than on fine-tuning approaches, believing superior context management and retrieval will deliver better results than model customization for enterprise code generation.

Labor market evolution

While senior engineers are favored in the short term, the long-term outlook favors junior engineers who can effectively leverage AI tools, shifting value from raw coding ability to architectural oversight and AI orchestration skills.