You have 1 article left to read this month before you need to register a free LeadDev.com account.
Estimated reading time: 6 minutes
Key takeaways
- AI-coding won’t be optional for long. Linear standardized on AI agents in a week and never looked back – 30% more PRs, 33% more issues closed.
- The new bottleneck is context, not code: Linear cut token usage 40% through context engineering and started writing more RFCs, not fewer.
- Keep humans in the loop until the infrastructure to remove them safely exists.
Many developers are now testing or actively using AI-coding agents. The outcome is clear: once you make the transition, the old software development lifecycle seems like an ancient way of working.
One group of engineers going “all in” on AI-coding agents are the folks at Linear, a product development system. “Compared to where we were a year ago, large language models (LLMs) are integral to almost every step of this now,” says Tom Moor, head of engineering, referring to their use in triage, bug investigation, and PR generation workflows. “When we’ve taken that away accidentally from people, they scream.”
In addition to building agentic capabilities into their product, Linear’s software engineers are relying on LLM-powered AI agents in their internal software development process to generate code, triage and fix bugs, mass assign issues, and more. It’s been especially useful for pre-investigation and pull request (PR) support when triaging incoming issues, he says.
In Linear’s triage workflow, Moor says over 50% of PRs generated by the coding agent are mergeable, making the coding agent one of Linear’s top contributors. That’s quickly becoming the norm for fast-moving engineering teams. “I don’t think it’s an unusual story from companies on the bleeding edge.”
However, this newfound dependency doesn’t mean engineering atrophy. Arguably, the dynamic has merely shifted. The new focus is on scaling agentic workflows. As Moor describes, this involves making tooling choices, optimizing context, adding guardrails, and keeping an eye on costs.
Your inbox, upgraded.
Receive weekly engineering insights to level up your leadership approach.
Getting on board with AI-coding agents
Up until recently, Linear adopted a bring-your-own attitude to coding tools. Engineers used a variety of methods, ranging from manual coding in VS Code, to AI assistants like Cursor and other LLM-based agents.
That was until January 7, 2026. On that day, Jori Lallo, co-founder of Linear, shared this internal memo:
“I’m asking everyone to install either Claude Code or Opencode and start writing code with Opus instead of by hand. Use IDE to review the diffs but rely on the model. Build a feeling for it.”
The month after this push, Linear tracked a notable increase in per-engineer productivity. In February, they averaged 52.7 PRs per author and 44.4 issues per author, representing a 30% and 33% increase from January, respectively. In March, those numbers climbed again, to 59.8 PRs per author and 55.2 issues per author.
“It’s a tough sell to get engineers to use a specific tool since everyone has their own specific ways that they like to work,” admits Moor. “But after that week, most people didn’t go back.”
Moving more engineers toward agent-forward coding tools has improved productivity. However, as within many other companies, it’s also introducing more code to review. “It’s already kind of a cliche to say that code review is the new bottleneck, but I’m sure getting that,” says Moor.
To avoid accumulating technical debt, Moor believes agents will also have a role in reviewing the rising amount of code. “I’m very bullish on AI, and with every problem it causes, the solution is more AI.”
Managing context bloat
Although coding agents are maturing quickly, their context windows still have a ceiling, limiting the data, code, memory, and Model Context Protocol (MCP) capabilities they can retain in a session. This has made context engineering, the act of fine-tuning what an agent is exposed to, an art in its own right.
At Linear, one way they’re slimming unnecessary bloat is by using rtk, an open-source tool that can be installed into Claude Code. “That’s been a cool find,” says Moor. The helpful command-line utility proxies common developer commands and reduces token usage by up to 90%, according to its designers.
For instance, using rtk, a Linear product engineer can send an issue to a local agent but massage the context to include only important details, including high-level IDs, and nothing more, says Moor. For Linear, such strategies have reduced token use by roughly 40%.
Since the barrier to generating code is so low nowadays, some engineering rigor has shifted to systems thinking and architectural schematics. As such, another tactic to optimize results with agentic AI is proper requirements gathering and upfront planning.
“We are writing a lot more requests for comments (RFCs) recently, and trying to make sure that we do have things hammered out,” says Moor. “That’s particularly useful once you throw that RFC as context in for the agent as well.”
More like this
Keeping the budget in check
Cutting-edge commercial models from AI companies like Anthropic or OpenAI grant the most power, and they’re intuitive to work with. However, as these companies introduce metered billing and increase per-token prices, potential cost issues are looming.
For Moor, this was bound to happen. The strategy was inevitable: get everyone addicted, then jack up the prices.
“Prices are gradually inching upwards rather than going downwards,” he says. In Moor’s experience, some lower-tier models are being phased out and replaced by newer cheap models that still cost more per token than the previous generation.
One tactic Linear is looking into to hedge against price hikes is by using free-to-use open-source models, such as Kimi, or OpenAI’s open weight models, for functions that don’t require state-of-the-art intelligence.
“We have LLMs all over our product, for doing bits of analysis like summarizing comment threads, helping generate notes, or reading documents,” says Moor. “You don’t need state-of-the-art models for that. Open-source models are perfectly fine.”
Other situations are not so clearcut. Moor shares one case, in which an engineer racked up $6,000 in Cursor bills while conducting a big migration from one library to another. The colleague didn’t want any visual regressions, and was automatically taking screenshots of each build to feed into each new prompt, consuming a lot of tokens in the process.
Even considering such one-off situations, AI use has remained within budgetary constraints at Linear. “It’s not an active day-to-day worry, but something we’re keeping an eye on,” says Moor. They’re a team of 45 engineers, and not breaking the bank on AI spending.
That said, things could change. For enterprises managing thousands of engineers, cost overruns with AI could amount to millions of dollars per month, he adds. As such, it’ll be worthwhile to track use and investigate cost optimizations, especially as the industry matures and companies operate agents at a larger scale.
Human in the loop, for now
Lastly, given the non-deterministic nature of LLMs and their ability to hallucinate, using them in a production software engineering context still requires proper guardrails.
Linear isn’t blindly accepting all AI-generated changes, and human engineers still must approve each eventual merge. They’re also adopting human approval gates for untrusted input, including a confirmation step to check that there isn’t an obvious prompt injection attempt before delegating work to an agent.
“Right now, this is still human in the loop,” says Moor. “I don’t know if that will last forever, but I would never say never on it being totally automated.” The agility gains of fully autonomous deployments might one day become a competitive advantage, but we’re not there yet, he adds.
New York • September 15 & 16, 2026
Cut through the hype.
Find what works at LDX3 New York
Building the system for agentic automation
At Linear, engineers are constantly polishing and refining small pieces all day long. The average engineer is merging six PRs a day, meaning they’re shipping hundreds and hundreds of times a day company-wide. “No change is too small,” says Moor.
Although there may be some reticence to adopt agentic AI due to security concerns, inserting little gates and points of approval throughout the software development lifecycle can help close these gaps.
While rising reliance on third-party agentic coding tools is becoming a big external dependency for programming at large, the productivity gains seem to outweigh the risks. This is convincing engineering leaders like Moor to keep building a feel for it, and maturing the practices and infrastructure to support it.
“You have to get ahead of automating as much of this stuff as possible.” The goal, he adds, is to “build the system to enable that.”
