Two Engineers, One OS, a Swarm of Agents

Inside the Swarm: How Sentient OS Builds With Parallel Agents

Two engineers. An operating system. A fleet of AI agents writing code in parallel across isolated git worktrees, each spinning up its own running instance of the product. This is how Sentient OS ships.

The Y Combinator-backed startup, founded in 2026 by Jesai Tarun and Aditya Vellanki, builds an on-device AI assistant for macOS that runs a local Gemma 4 E4B model on Apple Silicon to continuously process user context and proactively execute tasks. But the product is only half the story. The other half — the one rewriting how software gets made — is the development workflow itself. As their internship posting states: "We code with agents: parallel Claude Code and Codex sessions across git worktrees, with per-feature docs written so humans and agents can both load context in minutes."

The architecture solves a problem that has plagued AI-assisted coding since the tools arrived: two agents in one working directory overwrite each other. One refactors the API layer, another updates the frontend, both rewrite package.json, and the loser's changes vanish without an error message, discovered only when tests fail an hour later. Git worktrees fix this at the file system level. Each session gets its own directory on its own branch, sharing a single .git database, so parallel agents never touch the same file on disk. As the Claude Code documentation puts it: "Worktrees are one of several ways to run Claude in parallel. They isolate file edits."

Sentient OS takes this further. They create one worktree per feature, each with its own running application and URL. A parallel orchestrator spawns one Claude Code agent per worktree simultaneously. Per-feature documentation — written so both humans and agents can load context — travels with each worktree. As one practitioner of this workflow described the loop: "I walk away, agents work, I come back to review and merge." No stashing. No context switching. No "wait what was I doing."

The mechanics are precise. Create the worktree, run lint and test immediately, confirm the suite passes green, then hand the worktree to the agent. Each agent runs with a narrow, focused prompt. Each worktree produces its own diff. Review them one at a time and merge the ones you want. Isolation is what makes reviewing several parallel changes manageable. The ceiling becomes your review capacity, not your machine; practitioners report the sweet spot at three to four concurrent sessions; beyond six, review becomes the whole job.

Claude Code and Codex handle worktree creation differently. Claude Code's --worktree flag creates .claude/worktrees/<name>/ on a branch named worktree-<name> and starts the session there. Codex uses git worktree add plus --cd to start a session in a directory you created. Both approaches share the same repository history and remote as the main checkout. Subagents can run in their own worktrees too; add isolation: worktree to a custom subagent's frontmatter and delegated tasks run in their own checkout. Cross-session messaging lets agents pass findings between worktrees without human alt-tabbing.

The workflow has sharp edges. Worktrees stop two agents writing the same file. They do nothing about two agents changing the same behavior from opposite ends of the codebase, changes that merge cleanly and break at runtime. Shared runtime state (ports, databases, caches) must be provisioned per worktree. Lockfile churn from two agents adding dependencies produces merge conflicts miserable to resolve by hand. And a worktree tells you which files changed, not which commands the agent ran, what it tried and reverted, or why it made a decision.

Sentient OS's internship posting makes the hiring implication explicit: they want engineers who can operate this system. "You should naturally be intuitive with coding agents, and you should read and understand everything they produce. Our bar for quality and attention to detail is incredibly high." The role shifts from writing code to orchestrating agents: defining feature boundaries, writing the per-feature docs that load context for both humans and agents, reviewing diffs, and merging. "You code with AI tools ~daily. You have an intuition for their flaws and how to harness them well." This is the orchestration skill codified as a hiring criterion: not familiarity with Copilot, but daily, native fluency with the specific failure modes of Claude Code and Codex, and the judgment to know when to let an agent run and when to intervene. The two-person team ships an OS because they've built a development environment where the bottleneck is human review speed, not human typing speed.

The Product and the People: A Native Swift App Built by Young Founders

Sentient OS is not a wrapper. It is a native Swift application that wakes a Mac at 3 AM, processes the user's entire digital life — emails, messages, files, Granola transcripts, screenshots — entirely on Apple Silicon, and distills it into a searchable knowledge base written in plain markdown. The on-device model is Gemma 4 E4B, but the inference engine is a custom LiteRT-LM fork the founders built themselves: KV-cache reuse, speculative decoding, flash attention, and custom k-quants that let the system run nightly on 8 GB Macs. The user's hardware shoulders roughly 90 percent of the compute; their own ChatGPT or Claude subscription handles the rest. Sentient's servers store nothing. Marginal cost per user rounds to zero.

The product soft-launched in late July 2026 via a single Reddit post. Two thousand users arrived in 48 hours with zero marketing spend. The alpha processes around 3,000 screenshots overnight on a six-year-old iPhone; a modern phone chews through 6,000 in one session. If a user has 30,000 screenshots, the system works backward from newest to oldest so value appears after the first cycle. A "Sidekick" feature, computer use driven through the plain Codex CLI, executes prepared actions only when the user clicks. Nothing fires autonomously.

Behind the product are two founders who met the "shipping since childhood" bar before they could vote. Jesai Tarun, CEO, was a top-12 AI developer on GitHub at 18. He hacked iPadOS onto an iPhone 17 Pro via a bootrom exploit (1 million views), built Writing Tools — an open-source writing assistant that beat Apple Intelligence to market with 30,000 users and 2,400 stars — and has researched Apple's neural accelerators since high school. He is a CS student at UMass Amherst. Aditya Vellanki, co-founder, shipped consumer products at 12, built decentralized protocols at 14, and spent the last two years on enterprise systems. Aditya is 18; Jesai is 20. The company incorporated in 2026, employs two people, and operates out of San Francisco.

Y Combinator admitted them in the Fall 2026 batch with Aaron Epstein as primary partner. The pre-seed round totals roughly $1 million from YC, Afore Capital, and the a16z Speedrun scout fund. The YC listing describes the mission bluntly: "Nobody's built truly proactive AI, because it requires running inference on your entire life. In the cloud, that's insanely expensive and a privacy nightmare. But not on your own chip."

The open-source stance is deliberate, not performative. Tarun has stated he is considering an AGPL license so the code remains free for consumers while preventing competitors from commercializing his work. "Anyone can verify privacy claims by monitoring network requests; you would never have to take my word for it," he wrote on Hacker News. The first 250 early testers receive lifetime free access. Monetization, when it arrives, will be a roughly $2 monthly subscription for analyzing more than the last six months of data, with a one-time lifetime license option. The enterprise tier — the same engine plugged into Slack, Granola, Linear, and Notion — becomes the revenue engine.

This is not a research project. It is a shipping product built by two young founders who have been reverse-engineering Apple's ML stack since before they could legally drink, backed by the most selective accelerator in tech, giving away the consumer layer because the marginal cost is zero and the enterprise layer pays.

The Orchestrator's Shift: From Writing Code to Directing Agents

The Sentient OS founders code with agents daily. Two engineers, Jesai Tarun and Aditya Vellanki, spend their days decomposing features into specifications, spinning up such sessions across git worktrees, and reviewing the pull requests those agents open. The per-feature documentation they maintain serves as context loading for both humans and agents. This is what "orchestration" looks like in practice: you define the problem, you structure the environment, you evaluate the output, and you decide what ships.

The shift is measurable. Anthropic's 2026 Agentic Coding Trends Report found engineers report a net decrease in time spent per task category but a much larger net increase in output volume — more features shipped, more bugs fixed, more experiments run. AI enables productivity primarily through greater output, not simply doing the same work faster. At TELUS, teams created over 13,000 custom AI solutions while shipping engineering code 30 percent faster, saving over 500,000 hours with an average of 40 minutes saved per AI interaction. One enterprise customer finished a project their CTO had estimated at four to eight months in just two weeks using Augment Code powered by Claude. At Rakuten, Claude Code implemented a complex activation vector extraction method in vLLM, a 12.5-million-line codebase, in seven hours of autonomous work, achieving 99.9 percent numerical accuracy.

Gartner predicts that by 2028, 90 percent of enterprise software engineers will use AI code assistants, up from less than 14 percent in early 2024. The developer's job shifts from implementation to orchestration: problem solving, system design, ensuring AI tools deliver high-quality outcomes. Deloitte expects AI to drive productivity gains of 30 to 35 percent across the software development lifecycle. The primary human role becomes orchestrating AI agents that write code, evaluating their output, providing strategic direction, and ensuring the system solves the right problems for the right stakeholders.

This creates a new skill hierarchy. The most valuable skill is no longer writing code efficiently but articulating tasks with sufficient clarity, context, and constraints that agents can execute them correctly — what researchers call "intent engineering." Engineers who master orchestration can shepherd multiple features through development simultaneously, applying judgment across a broader scope than individual implementation previously allowed. They become more "full-stack" in their capabilities rather than being replaced; AI fills knowledge gaps while humans provide oversight and direction. For mid- and senior-level developers, demand for intangible skills (customer experience, cross-functional engineering, systems thinking, cross-product management) is expected to grow.

Team structures are bending accordingly. Deloitte projects conventional teams will shift toward fewer entry-level developers and more mid-level and specialized professionals, with a broader supervisory span for managers. New roles are emerging: AI governance specialists, prompt engineers and context designers, AI-augmented user experience designers. The number of AI architect roles is expected to almost double, from 30 percent today to 58 percent in the next two years. Almost a third of CIOs say orchestrating fellow tech leaders is essential in the next 18 months.

The abstraction layer has moved up. Traditional SDLC stages remain, but agent-driven implementation, automated testing, and inline documentation collapse cycle time from weeks to hours. Most tactical work (writing, debugging, maintaining code) shifts to AI while engineers focus on architecture, system design, and strategic decisions about what to build. Human oversight shifts from reviewing everything to reviewing what matters: teams maintain quality and velocity by building intelligent systems that handle routine verification while escalating genuinely novel situations, boundary cases, and strategic decisions for human input.

But the benchmarks expose a gap. EvoClaw shows overall performance scores dropping from over 80 percent on isolated tasks to at most 38 percent in continuous settings, revealing agents' profound struggle with long-term maintenance and error propagation. Context drift erodes coherent understanding of system-wide invariants as codebases grow beyond effective context windows. A small error in an early commit cascades into compounding failures; agents lack robust mechanisms for detecting and recovering from these chains. They optimize for immediate task completion without modeling long-term maintenance costs. Automated testing remains incomplete; agents can pass tests while introducing subtle semantic errors that manifest only under novel inputs.

The industry response is converging on a "human-in-the-loop, agent-in-the-driver's-seat" posture: neither fully autonomous nor fully human-driven. Organizations are establishing clear agent permissions, code review requirements, automated testing, security scanning, audit trails, approved development environments, human approval for high-impact changes, and metrics for quality, productivity, and cost. Investing in observability infrastructure becomes critical: agent systems require fundamentally different monitoring than traditional software. The productivity multiplier for those who master agent orchestration will far exceed the traditional "10x engineer" benchmark — not through faster typing, but through the ability to coordinate swarms of agents toward complex outcomes. The future developer may write fewer lines of code but make far more important decisions. AI becomes the hands. Developers remain the judgment.

Spreading the Template: How AI-Native Startups Are Rewriting the Playbook

Sentient OS didn't invent the parallel-agent workflow; it just pushed it further than most. Across the startup layer, the same pattern is hardening into a playbook. Lovable, a Swedish AI-native builder, reached within six months of launch and now serves over one million monthly users building software products, claiming 20x faster development than manual coding. Doctolib, a European healthcare platform, reports engineering ships features 40 percent faster while maintaining code quality. N26, the German neobank, achieved 70 percent automation across targeted processes within a year, deploying 15-plus AI applications for customer service and fraud analysis with implementation-to-testing cycles of one to two weeks. These aren't pilots. They're production metrics.

The data backs the anecdotes. Nearly 90 percent of organizations surveyed use AI to assist with coding today, and 80 percent report measurable economic returns: actual ROI, not projected value. More than half (57 percent) now deploy agents for multi-stage workflows, with 16 percent running cross-functional processes spanning multiple teams. In 2026, 81 percent plan to tackle more complex use cases: 39 percent developing agents for multi-step processes, 29 percent deploying them for cross-functional projects. Gartner predicts 40 percent of enterprise applications will integrate task-specific AI agents by year-end, up from under 5 percent. The shift from augmentation to automation is accelerating: directive conversations where users delegate complete tasks jumped from 27 percent to 39 percent over eight months, the first time automation exceeded augmentation.

But the industry reaction splits cleanly. Incumbents are layering agents onto existing workflows (what Deloitte calls "agent washing") and seeing poor ROI. Nearly half of organizations cite data searchability (48 percent) and reusability (47 percent) as blockers. Integration with legacy systems tops barriers at 46 percent, followed by data access and quality (42 percent) and implementation costs (43 percent). Gartner predicts over 40 percent of agentic AI projects will fail by 2027 because legacy systems can't support modern AI execution demands. Forrester warns at least one major enterprise breach will stem from agent misuse, not external hacking: a threat actor already used a jailbroken Claude Code to attack roughly 30 organizations in late 2025. PwC notes AI investments are increasingly judged on outcomes, not capabilities; agents that can't demonstrate clear ROI get paused or shut down.

The startups building natively avoid these traps. They don't retrofit. They design for agents from day one: git worktrees for parallel sessions, per-feature documentation that loads context for both humans and machines, orchestration over authorship. Token costs have dropped 280-fold in two years, yet some enterprises see monthly bills in the tens of millions, a signal that architecture matters more than access. Accenture put it bluntly: 2026 separates enterprises that deployed AI agents from those that transformed around them. The ROI ceiling isn't set by technology; it's set by the willingness to redistribute authority, redesign workflows, and trust intelligent systems with consequential decisions.

That trust is the real differentiator. The template spreading now isn't a toolchain. It's an organizational model: tiny teams, massive parallelism, documentation as infrastructure. Gartner predicts 80 percent of organizations will evolve large software engineering teams into smaller, AI-augmented teams by 2030. By then, AI agent-powered solutions could represent 60 percent of the total addressable software market. The race isn't between companies. It's between AI-native companies and incumbents still trying to bolt agents onto the old stack.

The Hiring Filter: What Sentient OS's Internship Posting Reveals

The internship posting for Sentient OS reads less like a traditional job description and more like a filter for a new kind of engineer. The role (Software Engineering Intern (Member of Technical Staff)) pays , remote within the US, with a hard visa requirement: US citizens or valid visa holders only. The stack is explicit: Swift, Windows, AI agents. But the real signal lives in the qualitative requirements, which map directly to how this two-person team actually builds.

The posting also reveals what AI-native startups don't prioritize. No mention of LeetCode patterns, system design interviews, or years of experience with a specific framework. Instead: "You understand systems deeply and love geeking out about interesting tech." "We'd love if you've built and shipped something real that people actually love." "You have an eye for great UX." The emphasis is on taste, shipping velocity, and systems thinking: the skills that survive when code generation becomes cheap.

Perhaps the most revealing line: "You don't mind spending 10x the time to perfect the last 10% (we pay you for all that 10x time :)." In a workflow where agents produce the first 90% in minutes, the human's value concentrates entirely in that final 10%: the polish, the edge cases, the UX decisions no model has context for. The posting makes explicit what the market is still pricing implicitly: the premium now attaches to judgment, not keystrokes.

This profile appears across the YC job board. Voiceops (W17) lists a "Founding AI Engineer" role; Vahan (S19) seeks a "Lead AI Engineer"; CoLoop (S21), Sully.ai (S21), Nextera Robotics (S20), Agave (W22), and Confido (S21) list various engineering roles in the same early-September window. The titles vary. The requirement doesn't: engineers who can orchestrate agents, evaluate output, and ship. The internship at Sentient OS is the entry-level version of that same archetype.

The compensation signal matters too. $3,000 monthly for a remote internship at a pre-seed company with a shipped product (2,000 users in 48 hours from one Reddit post) suggests the market values this skill set even at the intern tier. The team is two people. The intern works directly with the founders. There is no middle management to absorb the output. The feedback loop is immediate.

For candidates, the posting rewrites the preparation playbook. Contributing to open source matters more than coursework. Building a side project with agents (hitting their limits, learning their quirks) matters more than a high GPA. The "eye for great UX" requirement hints at a deeper shift: when agents handle implementation, the human's comparative advantage moves up the stack to product sense.

The visa restriction (US citizen/visa only) is a pragmatic constraint for a two-person team with no immigration bandwidth. But it also underscores how early this hiring pattern is. Larger AI-native companies sponsor globally. Sentient OS, at two people, cannot. The template spreads from the top down, but the earliest adopters hire where they can move fastest. The pair. The OS. The swarm. The loop closes.

Working in AI? Zero G Talent tracks the openings: see every open Databricks role, browse AI jobs, openings at Anthropic and Harvey AI, and the people building the field.