Written by Dmitrii Shilonosov, Senior Software Engineer at Bolt.

Like many things at Bolt, this project didn’t start with a formal requirements document or a top-down mandate. It started with a Slack message.

One of our colleagues pointed out a driver partner who had completed over 17,000 trips across five different profiles, yet the app was only showing the history of their latest profile ID. To riders, those 17,000 trips are a strong signal of experience and trust that we were simply leaving on the table. A casual observation like that might get lost in the scroll or stuck in a backlog for months.

At Bolt, the transition from a Slack ping to a Jira ticket happened in about 90 minutes. That ticket was the spark that BACA, the system we built in 2025, eventually picked up to turn a Slack conversation into a production-ready feature.

From prototyping to agentic engineering at scale

There’s growing excitement around AI-assisted code generation. Describe a feature to an LLM, get a block of code, and have a working prototype in minutes. That works well for personal projects or small tools. At Bolt's scale, it’s not enough.

Bolt has more than 600 engineers pushing code into its repositories every day, serving millions of users across the globe. Every line of code needs to be performant, thread-safe, and consistent with the architectural standards of the rest of the platform. A bidding service that manages thousands of concurrent requests cannot be built on quick prototypes alone.

The approach we employ is closer to what Andrej Karpathy calls agentic engineering: an autonomous agent produces a draft implementation, and engineers review, steer, and approve it. The agent writes, the human architects. We call this workflow intent-driven development. An engineer expresses intent through a ticket or a review comment, and the agent translates that intent into code.

To put this into practice, we're working on BACA (the Bolt Autonomous Coding Agent). It’s not a glorified autocomplete; it's a teammate that understands Bolt's specific server environment and helps scale engineering output without sacrificing quality.

A feature from start to finish

A common misconception is that AI agents are only good for scripts or isolated functions. We designed BACA to handle tasks that touch the core of Bolt's business. The driver-profile feature, for instance, required changes inside a high-traffic service responsible for matching drivers with orders. Adding even a few milliseconds of latency can have a massive ripple effect on the entire marketplace.

BACA started by taking the ticket and mapping out the dependencies. It identified the correct data structures and modified the service layer to aggregate driver statistics across profiles. It handled the manual work of service wiring and data mapping, the kind of digging that usually costs a developer hours of context-building.

The real efficiency, though, doesn’t come from the agent being perfect on the first try. It comes from how it handles review feedback.

When BACA submitted the initial draft, a human reviewer stepped in. The logic worked, but the implementation was not yet up to Bolt's production standards. The reviewer didn’t take over the task or rewrite the code. Instead, they left specific, high-level feedback directly in the pull-request comments, pointing out that the caching strategy needed to follow a more robust pattern and that certain fields must remain mandatory for consistency across the platform.

In a traditional workflow, the engineer would have to context-switch back, research the correct approach, refactor, and push an update. With BACA, the agent parsed the feedback, understood the requirement, and pushed a follow-up commit that reworked the caching layer and cleaned up code duplication from the first pass.

The automation didn’t stop at the application code. Separate pull requests were created to provision the required infrastructure changes, set up internal system access and credentials, and handle other supporting tasks.

From prototyping to agentic engineering at scale

There’s growing excitement around AI-assisted code generation. Describe a feature to an LLM, get a block of code, and have a working prototype in minutes. That works well for personal projects or small tools. At Bolt's scale, it’s not enough.

Bolt has more than 600 engineers pushing code into its repositories every day, serving millions of users across the globe. Every line of code needs to be performant, thread-safe, and consistent with the architectural standards of the rest of the platform. A bidding service that manages thousands of concurrent requests cannot be built on quick prototypes alone.

The approach we employ is closer to what Andrej Karpathy calls agentic engineering: an autonomous agent produces a draft implementation, and engineers review, steer, and approve it. The agent writes, the human architects. We call this workflow intent-driven development. An engineer expresses intent through a ticket or a review comment, and the agent translates that intent into code.

To put this into practice, we're working on BACA (the Bolt Autonomous Coding Agent). It’s not a glorified autocomplete; it's a teammate that understands Bolt's specific server environment and helps scale engineering output without sacrificing quality.

A feature from start to finish

A common misconception is that AI agents are only good for scripts or isolated functions. We designed BACA to handle tasks that touch the core of Bolt's business. The driver-profile feature, for instance, required changes inside a high-traffic service responsible for matching drivers with orders. Adding even a few milliseconds of latency can have a massive ripple effect on the entire marketplace.

BACA started by taking the ticket and mapping out the dependencies. It identified the correct data structures and modified the service layer to aggregate driver statistics across profiles. It handled the manual work of service wiring and data mapping, the kind of digging that usually costs a developer hours of context-building.

The real efficiency, though, doesn’t come from the agent being perfect on the first try. It comes from how it handles review feedback.

When BACA submitted the initial draft, a human reviewer stepped in. The logic worked, but the implementation was not yet up to Bolt's production standards. The reviewer didn’t take over the task or rewrite the code. Instead, they left specific, high-level feedback directly in the pull-request comments, pointing out that the caching strategy needed to follow a more robust pattern and that certain fields must remain mandatory for consistency across the platform.

In a traditional workflow, the engineer would have to context-switch back, research the correct approach, refactor, and push an update. With BACA, the agent parsed the feedback, understood the requirement, and pushed a follow-up commit that reworked the caching layer and cleaned up code duplication from the first pass.

The automation didn’t stop at the application code. Separate pull requests were created to provision the required infrastructure changes, set up internal system access and credentials, and handle other supporting tasks.

How the pipeline works

We designed the end-to-end flow to be fully automated from ticket to pull request:

(Optional) Some teams use automation that creates a Jira ticket directly from a Slack message.
An engineer labels a Jira ticket with todo.
The BACA pipeline detects the label, claims the ticket, and starts implementation.
The agent analyses the ticket, explores the relevant repositories, and produces draft pull requests with AI-generated descriptions.
Engineers review the PR and leave feedback by mentioning the BACA GitHub bot handle in comments.
The agent reads the feedback, pushes follow-up commits, and iterates until the reviewers approve.
A builder step validates that all required CI checks pass before the PR is marked as ready.

No manual triggers. No copy-pasting of prompts. The engineer's job shifts to what matters most: designing the solution, reviewing the agent's output for architectural correctness, and ensuring it meets Bolt's production standards.

How we ensure quality

A general-purpose coding agent wouldn’t survive Bolt's codebase. The platform has its own conventions, architectural patterns, and quality standards that every change must follow. When designing BACA, my team and I closed that gap with three layers of customisation:

Strict rule sets. The agent operates under a defined set of rules that encode Bolt's coding standards. It doesn’t invent its own style; it follows Bolt's.
Code generators are exposed as standard tools. We built specialised code generators and made them available to the agent through MCP (Model Context Protocol) — the open standard for connecting AI agents to external capabilities. The agent calls these generators the same way an engineer would. Because the generators produce output that already conforms to Bolt's patterns, the agent's code is locked to the correct style from the start. It builds logical blocks using our generators and writes only the connecting logic itself.
Automated review agents. Before a human reviewer sees the pull request, a separate review agent checks that the output follows Bolt's code style, naming conventions, and structural rules. Violations are automatically detected and fed back to the coding agent.

One of the first things I insisted on when designing the pipeline was that no agent-generated code should reach a human reviewer without passing automated checks first. On top of the layers above, several safety guardrails run before a human even opens the pull request:

Automated sanity checks. An automated validation step checks the agent's output against structural rules before the PR is created, using the same checks that run for any other PR, no matter who created it.
CI/CD pipelines. Standard test suites and linters run against the agent's code exactly as they would against any human-written code.
Mandatory human review. Every BACA-generated pull request requires at least two human approvals before it can be merged. The agent cannot merge its own work.

These guardrails mean the agent operates within a controlled corridor. It can produce code quickly, but it cannot ship anything without passing the same bar Bolt sets for every engineer.

Results so far

By using agents this way, we’ve fundamentally changed the role of our developers. We’re moving towards a model where our engineers act as architects and guardians of quality, while the agents handle the transcription and boilerplate.

In the last two weeks alone, BACA opened 239 pull requests across Bolt's repositories — roughly 18% of all pull requests created in that period. Of 72 merged agent-generated pull requests, 32 were merged without any human code modifications — the agent's output passed review as-is.

In the remaining 40, human reviewers made targeted adjustments, steering the agent's draft toward production standards. Across those merged pull requests, the agent produced more than 40,000 lines of code changes, giving engineers a ready-made starting point that they could refine rather than build from scratch.

What comes next

When hundreds of people contribute to one codebase, consistency is everything. Humans get tired; they might forget to use the standard caching library or skip a mandatory field check when they’re in a rush. An agent like BACA doesn’t get tired. It follows the patterns it’s taught, and when corrected, it applies that correction across the entire pull request immediately.

Today, BACA handles the coding and review cycle. Looking ahead, we’re working towards a full feature-development lifecycle where BACA clarifies requirements with the ticket author, reviews the refined scope with product managers, and then develops an implementation plan that engineers approve before any code is written.

Beyond engineering, we see BACA enabling support teams to use the agent for investigation and diagnostics, turning it from a coding tool into a broader operational partner across Bolt.

Join Bolt

Grow fast, make a global impact, and help make cities for people, not cars. Explore hundreds of open roles on our careers page.

Transforming Slack conversations into production-ready features with BACA at Bolt

We’re hiring!

Recent posts

AI-prototyping: Reality check