3.1 Extreme Parallelism: Phone + Hundreds of Agents#

Boris’s way of working is staggering:

• His primary work device is an iPhone, orchestrating everything through the iOS Claude app
• Launches 10–150 PRs per day, with a single-day record of 150
• Runs hundreds of AI Agents in parallel simultaneously, with thousands executing deep tasks each night
• Opens 5 tabs in the local terminal at once, plus 5–10 sessions on the web, taking turns launching Claude Code across environments

His Ansible and Nginx configs are all Claude-generated. “Since last year, I can’t understand any of this anymore.”

3.2 Loop Mode: His Favorite Feature#

Loop is a cron-based scheduled agent orchestration mechanism. Boris has deployed dozens of long-running Loop instances:

• PR babysitting: automatically fixes CI conflicts, automatically rebases
• CI health maintenance: automatically fixes flaky tests
• Social listening: scrapes Twitter feedback every 30 minutes and automatically clusters and organizes it

1
#Check CI status every 30 minutes and fix failing tests
2
/loop 30m check the latest CI run; if there is a flaky test, fix it and open a PR

3.3 Plan First, Then Execute#

80% of Boris’s sessions start with Plan Mode (enter by quickly pressing Shift+Tab twice):

1. Communicate requirements back and forth with Claude until a satisfactory plan is generated
2. Switch to auto-execute mode and let Claude make all the code changes
3. Review the results and iterate if necessary

However, he also mentions that as model capabilities improve, “a month from now he won’t need plan mode anymore”—the model will plan on its own during its thinking process.

3.4 Multi-Agent Collaborative Communication#

The most mind-bending point: his Claude will message colleagues directly via Slack.

For example, if it sees that someone in git blame recently modified related code, Claude will automatically send a Slack message asking clarifying questions. Agents can autonomously team up, claim tasks, and divide labor. Boris believes this is the core capability of the next generation of programming tools—“letting multiple AIs work like a team.”

3.5 Shared Memory: CLAUDE.md#

The team maintains a shared CLAUDE.md in the Git repository, recording development commands, code style preferences, project architecture conventions, and so on.

The key mechanism is: whenever Claude is found to have made a mistake, @Claude in the PR comment and have it automatically update CLAUDE.md. This way, a mistake is only made once, and knowledge continuously accumulates within the team.

3.6 Routines: They Run Even Offline#

Routines are server-side persistent routines that can keep running even when the device is offline. This means that even when Boris is asleep or on a plane, his Agents are still working—when he wakes up in the morning, the PRs are already waiting for review.

5.1 Build Products for the Model Six Months From Now#

We don’t build products for today’s model, we build products for the model six months from now.

This is Boris’s most core product philosophy. He refuses to patch current model deficiencies—“patches directly become negative assets once the new model comes out.” Rather than spending time fixing the current model’s limitations, it’s better to focus on a product framework that remains competitive once model capabilities overflow.

5.2 Model vs Orchestration Layer#

The weighting between the model and product design is about 50/50, but in the long run, improvements in the model’s own capabilities will reduce the necessity of the orchestration layer.

He even predicts that Claude Code may be only 100 lines of code a year from now—because the model will increasingly autonomously do the right thing, without needing complex external logic to guide it.

Two Types of Weakened Moats#

1. Switching costs: cross-platform migration will be automated, and lock-in effects will weaken
2. Process advantages: AI can reproduce and optimize complex business processes, and once-”proprietary operations” are no longer scarce

Disruption in the Next Decade#

• The number of disruptive software companies in the next decade will be 10 times that of the past decade
• Small teams, with the help of AI, can go head-to-head with big companies
• The title “software engineer” may disappear, replaced by Builder: defining system goals, reviewing output, and deciding what problem to solve

A Fundamental Shift in the Programming Paradigm#

From “humans write code, AI assists” to “AI writes code, humans make decisions.” A developer’s core ability is no longer syntax proficiency, but problem-definition ability, system design taste, and judgment.