Intercom ships fast. It’s the leading AI customer service company that delivers Fin, the best performing AI Agent for customer service, and a next-generation helpdesk for human support. Their infrastructure and product teams manage large Ruby and JavaScript codebases, distributed systems, and a constant stream of high-impact changes. But as senior engineers explored agent-driven development, they found existing tools struggled with context, consistency, and reliability. After trying Cursor and multiple LLM coding tools, Intercom adopted Augment Code to enable parallel agent workflows, accelerate large feature builds, and give engineers a trusted way to offload the repetitive details of coding—without sacrificing quality.
Intercom's complex systems
Intercom engineers work on complex systems where accuracy matters including:
- Large, interconnected Ruby codebases
- Infrastructure changes that require precise, multi-file edits
- Fast shipping culture with high technical standards
- High variance in team workflows and tooling preferences
For Intercom’s Senior Product Engineer Marvin Bitterlich, earlier AI coding tools weren’t reliable enough:
“They hallucinated, forgot files, or just lost the plot. I didn’t use AI at all until agentic workflows arrived—then everything changed.”
To make agents practical, Intercom needed one thing above all: reliable context management.
Why choose Augment Code?
Previously, Marvin evaluated Cursor’s agent mode first—and hit immediate limits. Forgetting files, inconsistent plans, and poor context handling made real engineering workflows impossible.
Augment stood out immediately for its stability and context fidelity.
| Requirement | Augment advantage | Marvin’s take |
|---|---|---|
| Context accuracy | Deep codebase understanding across large repos | “It just knows where things are. I speak naturally—Augment finds the right files.” |
| Consistent task execution | Sequential Thinking + stable agent plans | “Other tools lose track. Augment actually finishes the work.” |
| Scalable agent workflows | Local + remote agents with GitHub Issue/PR integration | “I ran hundreds of PRs a week across coordinated agents.” |
| Large-system onboarding | Accurate summarization + diagram generation | “I read far less code myself—Augment connects the dots reliably.” |
| Workflow flexibility | Works across VS Code, JetBrains, CLI | "I switch between VS Code and CLI workflows, and Augment works well in both." |
Showing context engine tool call
The tipping point: Augment succeeded on both small prototypes and massive feature builds where other tools failed.
“If I have to work in Claude Code, I still always have the Augment Context Engine MCP installed. ”
The results:
Parallel agent development at scale
Marvin used Augment’s local and remote agents to orchestrate complex multi-PR workflows:
- Plan a feature
- Split into GitHub issues
- Launch remote agents to generate draft PRs in parallel
Remote agent view showing multiple remote agents drafting PRs
- Review and merge
- Have Augment reconcile the merged work
“We built a large internal app—over 100,000 lines of code. I merged 200 PRs in a week. Augment got tasks done, and I reviewed every line.”
High-quality code, consistently delivered
90–95% of Marvin’s production code over the last months was written by Augment—at the same quality bar Marvin sets for himself.
“It’s the same code I’d write. I just don’t have to type it.”
Faster onboarding into unfamiliar systems
Augment summarizes unfamiliar services, traces dependencies, and generates mermaid diagrams for team discussions.
Engineers read only the critical files—Augment handles the rest.
More focused engineering hours
By offloading detail work, Marvin stays less fatigued and more productive:
“I can work for 12–14 hours on a weekend project because Augment handles the exhausting parts. I just review and guide.”
Intercom engineers now waste less focused time on boilerplate and more on solving real problems.
A tool for experienced engineers, not vibe coding
Marvin notes that Augment succeeds where “vibe coding” fails:
“You still need to understand the problem. I solve it in my head—Augment handles the implementation.”
Marvin’s advice:
- Frame the problem clearly (like talking to a colleague)
- Give high-level intent, not long prompts
- Only ask the AI to do tasks that are possible and well-understood
- Use agents for parallelizable work; use chat for precision edits
The result: true engineering acceleration, not shortcut-driven churn.