Video · 2026-03-10 · 1h 16m · 6 moments

From Code Generation to Full System Autonomy: Blitzy's Approach

✦ AI generated

timeline · colored by role

Mechanism◆

Dynamically recruiting swarms of tens of thousands of agents and using the database itself as the orchestration layer—rather than relying on a single central orchestrator—lets Blitzy reliably write hundreds of thousands to millions of lines of code that compile, pass all tests, and render pixel-perfect UI.

Sidhant describes Blitzy's core architecture—agent swarms coordinated via the database rather than a single orchestrator—as the key to scaling autonomous code generation to millions of lines that compile and pass all tests.

transcript

Sidhant Pardeshi: The approach that we took has been to dynamically recruit multiple swarms of agents and use the database as part of the orchestration layer. And you can recruit tens of thousands of agents, but not have to worry about this single orchestrator that's keeping track of everything that's happening.

Claim◆

Generating code with AI has become a commodity and is now easy; the real, much harder challenge is producing code that meets an organization's standards, is secure, and is genuinely production-ready enough to be accepted.

Sidhant argues that while AI can write code easily, the true bottleneck in autonomous development is getting that code accepted—meeting standards, security, and production requirements.

transcript

Sidhant Pardeshi: You can write a lot of code uh and code is a commodity now. Like getting AI to write code is is very easy. Getting any code is easy. Getting code that follows your standards, codes that code that is really good, uh code that is secure um code that is ready for production is a completely different story, right?

Example

Applying a naive 'Ralph Wiggum loop' (repeatedly re-running the same instruction) on tools not designed for it fails badly at scale, as shown by an autonomously-built C compiler project where a basic hello-world program didn't even compile.

Sidhant cites an autonomously-built C compiler project—where hello world didn't even compile—as evidence that simply looping an agent on a spec ('Ralph Wiggum loop') fails without real context and agent engineering.

transcript

Sidhant Pardeshi: you had uh for example, Anthropic put out this huge project that was a C compiler. Um and the very first issue, not very first, but one of the most popular issue on that forum is that this repo should not exist because a hello world does not compile on this compiler, right?

gives example · 1

Mechanism

Building a hybrid graph-and-vector representation of an entire codebase—mapping relationships and semantic summaries during ingestion—lets agents jump instantly between distant points in a 10-million-line codebase instead of burning tokens traversing files via grep alone.

Sidhant explains that Blitzy built a hybrid graph-and-vector map of the entire codebase as an 'anchor point,' letting agents traverse huge codebases instantly rather than burning tokens searching file by file with grep.

transcript

Sidhant Pardeshi: we've built a hybrid between a graph and a vector, um, that where you have this ingestion process with let's say for example that where it understands the entire code base, maps out the relationships, does semantic summarization and aggregation. And now you have this map of the entire code base.

Data

Even as nominal model context windows have grown from about 10,000 tokens to 1 million tokens, the effective context window—where models actually perform reliably—has been stuck around 80,000-120,000 tokens for roughly two years and isn't likely to change soon.

Sidhant points out that despite headline context windows expanding to a million tokens, models' effective working context has plateaued at roughly 80-120k tokens for about two years, a limitation he doesn't expect to change soon.

transcript

Sidhant Pardeshi: the effective you you've gone from 10,000 tokens to 1 million tokens and you've gone from, uh, you know, maybe 10,000 to 200k tokens and then 1 million, but we've been stuck at 80k to 100k tokens, um, 80k to 120, I would say, with the latest models, since 2 years.

explains mechanism · 1

Example

Text-based agents.md instruction files work fine for smaller codebases (under roughly 70,000-100,000 lines) but cannot generalize across large enterprise codebases and can produce contradictory instructions that leave agents to guess, as happened when a rule to use fakes instead of mocks conflicted with an existing codebase pattern of using mocks.

Sidhant argues agents.md files break down at enterprise scale, describing an internal case where a rule to use fakes instead of mocks conflicted with existing code patterns, leaving the agent to inconsistently pick one or the other.

transcript

Sidhant Pardeshi: we have a rule that says in in Python, uh only use fakes and not mocks for writing tests, all right? Like you can think of that as a agents.md. Um but then in the code base, we've extensively used mocks, right? And we have another instruction that says always mimic the patterns that we've already used in the code base, right? Now, what do you expect the agent to do, right?

Highlight slides

Database-as-orchestrator scales agent swarms✦ from: Dynamically recruiting swarms of tens of thousands of agents and using the database itself as the orchestration layer—rather than relying on a single central orchestrator—lets Blitzy reliably write hundreds of thousands to millions of lines of code that compile, pass all tests, and render pixel-perfect UI.Output: production-grade code at unprecedented scale✦ from: Dynamically recruiting swarms of tens of thousands of agents and using the database itself as the orchestration layer—rather than relying on a single central orchestrator—lets Blitzy reliably write hundreds of thousands to millions of lines of code that compile, pass all tests, and render pixel-perfect UI.AI Code Generation Is a Commodity✦ from: Generating code with AI has become a commodity and is now easy; the real, much harder challenge is producing code that meets an organization's standards, is secure, and is genuinely production-ready enough to be accepted.The Real Challenge: Production-Ready Code✦ from: Generating code with AI has become a commodity and is now easy; the real, much harder challenge is producing code that meets an organization's standards, is secure, and is genuinely production-ready enough to be accepted.

Related episodes

Video · The Cognitive Revolution

Intelligence on the Edge: Liquid AI's Ramin Hasani on the Search for Device-Native Foundation Models

→

Video · Machine Learning Street Talk

The Dangerous Illusion of AI Coding? - Jeremy Howard

→

Article · The Pragmatic Engineer

Context engineering with Dex Horthy

→