At Parahelp, we are building products to “Increase happy customers on the internet”. As a first step, we’re building the best AI customer support agent for software companies.

Concretely, this means we get to work on solving a lot of exciting problems at the edge of what’s possible with llm agents. It also comes with a lot of infrastructure challenges as we’re rapidly scaling and helping companies like Perplexity, Cursor, ElevenLabs, Replit, and Framer and their hundreds of thousands of customers every day.

We started Parahelp about a year ago. The AI customer service space seemed crowded then, just as it does today, but when talking with software companies, we quickly realized that for complex ticket automation, there were no solutions that actually worked well for them - and that their customers therefore loved.

This is because automating a support ticket for a software company is often far more complex than simply relaying information from an automatically scraped FAQ page. In fact, the AI customer support agents our customers build with Parahelp often take >100 steps, using >10 tools across >5 emails back-and-forth to resolve a single ticket.

An example is a potential bug report. This usually involves first troubleshooting with the user based on a knowledge base, previous tickets, and relevant user data from internal tools. If this does not resolve the issue, the next step is to gather additional context. This happens by searching in Linear and even the codebase, while simultaneously asking the customer for more details, like screenshots and device details. Then, the agent needs to coordinate the engineering handoff and prioritization via Slack/Linear, and, lastly, make sure to follow up with affected users by proactively listening to events in external systems like Linear and GitHub.

And even with our current AI agents having the capabilities to automate tickets like the above, there are still many tickets they can’t solve yet, and that can easily take a human agent - often with technical experience - hours to solve.

This challenge is what we are excited about. It’s really fun to build at the frontier of what’s possible with “agentic systems” today, and even more so when every percentage point we improve our AI agents' ability to automate tickets translates into thousands more happy customers on the internet every day.

To make the best AI customer support agent, here are the two main areas we’re currently focused on and working on:

Agents that write code

We recently released a code-gen agent for building evaluation environments in real-time.

This agent has access to the same policies and tools as the company’s AI customer support agent. Then, using these and reading the code, it can write its own evals in code (yaml scaffolding + tool simulation). We’ve also given this agent its own version of “git for support policies” and set up sandbox-like infrastructure behind the scenes, allowing our customers to create “branches” with local versions of the company’s AI customer support agent while drafting new policies and tools.

This enables our customers to simply tell our agent about a new policy, and then it will draft it in a local filesystem, create a full evaluation environment to run sandboxed tests for the new policy, run the tests, reflect on test results, and finally optimize its own prompts based on the results. This loop then repeats until all tests pass, at which point our customers can merge the fully reliable, AI-agent-optimized policies into production.

The infrastructure that makes this possible is all inspired by leading code-gen agents, and it works surprisingly well for AI customer support agents! Building AI customer support agents with Parahelp now more closely resembles vibe-coding with Claude Code, in contrast to previous solutions where companies were required to write their own prompts, set up inflexible workflows in drag-and-drop interfaces, and manually QA these llm workflows.

And we’re still only scratching the surface of what we actually believe we can build by combining the models’ fast-improving code capabilities with sandbox access, where they can write code and policies in .py and .md files optimized for solving complex support tickets.

Agents that use the browser

While most benchmarks seem saturated with recent model releases, and the actual impact of improvements is therefore harder to feel, this is not the case with computer-use agents! Opus 4.5 jumped to 66.3% on OSWorld. This is an increase of more than 20 percentage points over Opus 4.1 in just 3.5 months!

We think the steep rate of improvement in the model’s ability to reliably navigate a computer is extremely exciting for AI customer support agents. The primary reason is that the biggest bottleneck in resolving complex tickets for our customers today is not intelligence, but rather not being able to access the same tools to gather necessary context or take actions that human agents can.

This is because every company has a range of internal tools, authentication systems, common SaaS apps configured in specific ways, etc., that they use to resolve tickets. All of this is straightforward to set up and use for a new human agent joining the team, but it takes a lot of effort to replicate for an llm agent! Every context or action tool requires engineers inside the company to set up the necessary API endpoint, authentication, and coordinate with the support team on what’s needed for the llm agent specifically. And not only does it take a lot of time and custom code for a company to replicate this for an llm agent today, but it’s also not generalizable, and it needs to be maintained continuously as a company’s tools evolve. Computer use agents solve this.

Currently, we’re exploring how to turn screen recordings into “computer use workflow agents” that can be tested thoroughly, how to do deterministic regex/code checks in the DOM on top of vision-based computer-use agents (e.g. you can only ever click “refund” button if the table-row contains “x@gmail.com” string), self-healing, authentication for agents to internal dashboards and through services like Okta, and much more.

—

If you’re also excited about code-gen, computer-use, or scaling infrastructure that millions of customers interact with, then we would love to meet you. Please email anker@parahelp.com with something cool you’ve built.