Every coding agent begins with the confidence of a new notebook and, too often, the institutional memory of a goldfish. You explain the architecture, establish the constraints, reject three tempting mistakes and finally reach a useful answer. Then you change tools. The reasoning remains behind, sealed in a chat history that your next agent cannot read and your teammate may never see. The code travels through Git. The story of the code takes the bus and misses its stop.
Narayana Aaditya Ganeshkumar - Nars online and to most people who follow his work - has made that gap his current problem. He is the co-founder and CEO of Skillsync, a San Francisco developer-tools startup in Y Combinator's Winter 2026 batch. Its desktop app moves sessions, memory and reusable skills across coding agents and teammates. Begin in Claude Code, continue in Codex, and preserve the discussion that got you there. The promise is plain enough to fit on a sticky note. The implications spread across the whole wall.
The premise is that a useful AI session contains more than output. It holds decisions, false starts, tool history and the little turns of judgment that distinguish a merely working patch from one a team actually wants. Lose those, and the next person inherits a result without its map. Engineering has always had this problem. AI has simply made the pile of missing maps taller, faster.
The interoperability habit
Skillsync looks like a very 2026 company, but its founder's recurring subject is older: systems that need to cooperate despite having different rules. Before Skillsync, Aaditya worked around Juspay's Hyperswitch, an open-source payments project written in Rust. Payment orchestration is an ungainly phrase for an elegant ambition. A merchant should be able to work across processors without rebuilding the machinery for every provider.
Aaditya wrote about that world in practical terms: routing, conversion, processor choice and the scarcity created by an abundance of options. He also co-authored an essay called “The Inverse Conway Maneuver” with Nishant Joshi, now his co-founder and Skillsync's CTO. It examined the feedback loop between an organization's communication and the software architecture it produces. Architecture shapes collaboration; collaboration returns the favor. The idea has followed both men into their next company.
The founders describe years connecting dozens of payment processors and growing their open-source Rust work past 30,000 GitHub stars. That experience matters because compatibility is never just a menu of import buttons. Different systems preserve different objects, assumptions and histories. Moving a transcript without losing its meaning is closer to translation than copying a file. Skillsync's translation engine, txcript, is open source and written in Rust - a small bridge with its beams exposed for inspection.
There is a pleasing continuity here. First, help payment systems exchange instructions. Then help coding systems exchange thought. The cargo has changed from transactions to context, which is the sort of career progression nobody explains at school because it sounds ridiculous until it happens.
“Software is a team sport.”Narayana Aaditya
From hidden talent to trapped context
Skillsync did not arrive at its present product in a straight line. Its earlier public launch focused on understanding engineering work itself. The product read pull requests, reviews and comments to create a living map of skills. Instead of treating GitHub as a field of green squares and contribution counts, it looked for the shape of a person's work: research, firefighting, collaboration, craft.
That version of the company fitted Aaditya's unusual combination of interests. He studied human resource management at XLRI in Jamshedpur from 2018 to 2020, then moved through consulting, product and open-source technology. He can talk about payment routes and organizational psychology without treating either as decorative vocabulary. When asked to classify engineers as good or bad, he resisted the premise. He wanted the product to find possibilities in different contributors and understand how they worked.
The thought is generous, but it is also operational. Teams routinely know less about their own abilities than they imagine. Expertise hides in a review comment. A rescue from six months ago is remembered only by the person who performed it. A careful workflow remains private because nobody turned it into a playbook. The first Skillsync tried to reveal that submerged work. The current Skillsync tries to keep today's thinking from becoming submerged tomorrow.
shows real ability
hold real reasoning
make it reusable
Products pivot. Convictions migrate. Across both versions, Aaditya keeps returning to the same overlooked asset: evidence of how people actually do the work. A repository is richer than a résumé. A session is richer than its final patch. A team's culture is richer than the checklist in its onboarding document. Skillsync's business has changed shape around the belief that this richness can be made visible and portable.
The jazz problem
Aaditya's most revealing metaphor is not technical. It is musical. He has compared good collaborative building to jazz, where players can respond to an unexpected phrase and resolve it together. Improvisation works because everybody is present, listening and sufficiently practiced to make surprise useful. Four musicians recording alone and mailing their tapes do not have the same conversation.
His critique of modern engineering follows from that picture. Teams discuss, write a specification, send agents into separate corners and meet again through tickets and pull requests. The hand-off delivers the code after the live reasoning has cooled. Teammates see what happened but not the moment when one approach suggested another, or when a bad idea revealed a better one.
Aaditya and Joshi say they work differently when a problem is hard. They enter the same session and build through it together. One chooses an approach the other would not have considered; seeing it develop changes the plan. Their Campfire concept turns that habit into a product: a shared live session teammates can join, observe and continue. The name is cozy. The ambition is architectural. It treats collaboration as part of the development environment, not the meeting scheduled around it.
The jazz comparison also tells us something about Aaditya's style. His public writing prefers concrete metaphors, lower-case launch notes and questions over decrees. He can be earnest about engineering empathy, then tweet a two-word maxim like “Show, don't tell.” It is a founder's voice tuned for conversation, not a white paper wearing expensive shoes.
Action as a research method
Asked what advice he would give his younger self, Aaditya offered a compact principle: action creates information. If he and Joshi lack clarity, he said, they usually have not spoken to customers enough. The cure is not another private romance with the idea. Put the work outside. Let another person misunderstand it, refuse it, use it strangely or point to the part that matters.
That principle helps explain the visible evolution of Skillsync. The company has been willing to ship different expressions of its thesis: skill graphs from open-source activity, developer search, team intelligence, session portability and shared building. Each launch leaves evidence. The founder who argues for reading real work instead of claims is applying the same standard to his own company. A shipped product is an argument that customers can answer.
It also offers a useful restraint on the grand language surrounding coding agents. Everybody is using them, Aaditya has observed, while the norms of good use remain unsettled. His Build Logs series responds by speaking with opinionated engineers about what they actually do, then turning those workflows into installable skills. Again: observe the practice, capture the context, make the tacit thing transferable.
There is no guarantee that one shared memory layer will become the common room for a fragmented tool market. Agent makers can add portability. Teams may prefer their own stacks. Formats will mutate with cheerful indifference to anyone building a translator. But those uncertainties sharpen the question rather than dissolve it. As generated code becomes abundant, what becomes scarce? Aaditya's answer is not tokens. It is accumulated judgment - the reason a team knows which code to keep.
The code is the artifact everyone can see. Context is the invisible apprenticeship behind it.
A company about remembering
Aaditya's path into this problem is almost suspiciously neat in retrospect. Study people and organizations. Work in consulting. Help build open-source infrastructure that reconciles incompatible payment systems. Notice how much engineering ability hides in public code. Start a company with a technical collaborator from that world. Follow the hidden work from repositories into AI sessions.
Of course, lives are only neat when arranged after the fact. The more honest version contains launches, revisions and the steady embarrassment of discovering what customers meant. Skillsync's current focus is young. Its team is tiny. Its subject changes weekly because the agent ecosystem seems to regard stability as a personal insult. Those conditions suit Aaditya's method: move, listen, collect information.
The aspiration underneath the product is stable enough. He wants software tools to bring out the best in engineering teams. That means preserving the clever fix, but also the standards, rejections and explanations around it. It means allowing a teammate to arrive in the middle of the song and hear the key. It means treating context not as exhaust from the real work, but as part of the work itself.
The industry has spent several years asking how much code an agent can produce. Aaditya is asking what happens after the production: who remembers, who learns and whether the next session begins any wiser. It is a less theatrical question. It may be the one teams keep.