There Is No Middle: Sidu Ponnappa on Agents, Debt, and the Human in the Loop

Subtext by Zerodha — Sidu Ponnappa with Pranav — “What is an AI-native firm?,” CPISElKbFFg

Give a median Indian developer an agent and the $30-an-hour band disappears. Sidu Ponnappa’s claim is not that juniors vanish. It is that “good enough” does. Either they produce slop or they produce bangers. There is no adequate-with-an-agent. That is the same pattern he sees in architecture: you can oneshot a website; you cannot nail three brand pillars and walk away. The agent will drift. A human has to protect the pegs. Miss one decision in a day of hundreds and the integrity of the system compounds into debt — not just technical anymore, just debt — until around two weeks the whole thing turns to slop at an architectural level.

He sold C42 to Gojek, ran Gojek engineering, and now runs Realfast, which he insists is a manpower company, not a SaaS. In venture that used to be the kiss of death. He thinks it is the only Lindy invariant left before the singularity: a human in the loop until there isn’t one, at which point looking further is pointless unless you are a frontier lab. He could be horrendously wrong. He is betting years and capital on it anyway.

Not an intern. System one.

The popular definition of an agent is agency: iteratively pursue a goal with tools. The interesting definition, he said, is the one people avoid because they anthropomorphize. From GPT-3 through 3.5, 4o, Sonnet, Opus, Astra, there is agency with subtle limits. A ThoughtWorks senior who went to Google told him an architect can hammer down three pegs, no more, and rotate the system around them. Agents will not honor those pegs unguided. Trivial systems of any kind — comms, software, sales, recruiting — still need a human protecting integrity. Without that, inductively, the system turns to slop. Bad copy is the trivial example. The insidious one is the missed architectural decision you don’t notice.

Real usage is not an external person running by itself. You use the agent as an extension of your subconscious: system two sets the objective, system one executes. Magnified system one. The intern-to-god maturity scale is therefore the wrong axis. The question is how deluded you are by your own subconscious, because that is what the agent magnifies as it tries to please you. Fail to run defense and output degrades. In a fragile mental state, he said this without lightness, you are looking at full-blown schizophrenic breakdown. Treat the thing as an infohazard of the highest order. The high is real. The game is craft: you are already a superhuman cyborg when you work with it, the same way intraocular lenses and knee replacements were. Six million lenses a year for decades. Superintelligence is two-sided, like the subconscious. How you use it in pursuit of craft is now the whole game.

He has wanted transhumanism since ninth standard, after Asimov’s “The Bicentennial Man” — a robot seeking the invariant of dying. Humans, he said, are walking the other road: hygiene, medication, computers, godhood. “Play God, pay the price” he calls Western rock thinking. Water thinking, the Tandava in the bones: you were born already paying the price. The singularity will shatter rocks that cannot bend. Multivariate Indian culture, he joked, is actually well suited to navigate it. Uncle thinking borrowed from the West masquerades as Indian thinking. The reality is older.

Alpha on a three-month clock

Businesses have always extracted alpha from regulation, politics, special economic zones. They have never had to do it on a three-month horizon. In the foothills of the singularity, that is the operating window. Pivot was normalized in startups because it is a bad thing; even in 2022 a quarterly pivot was sketch. Realfast inherited “build to a thesis” from the last cycle and lagged step-changes by six months. They spent four months solving diffs so model output could land in a running system. GPT-4o invalidated the premise. It took four to five months to retire. Sonnet 4 and Claude Code were the first “true agent”; before that, agent was a word you sold to people who didn’t know better, including very large 2024 enterprise platforms. They only internalized “revisit strategy on every step function” around Opus 4.5, October–November 2025. They had closed $100k that entire year. Deals they had been saying no to as off-thesis were now on-thesis. By January 20 they had seven-figure revenue. A 10x in two months because they pivoted.

The deals were not ones they should have taken earlier. They would have failed. The example: a 35–40-year-old Delphi system in regulated financial and corporate-secretarial work — ESOPs, you cannot mess around — plus a .NET rewrite, two stages of legacy both in flight, a hard deadline under three months. On 4o they would have crashed. Sonnet 4 and Claude Code made it risky. Opus 4.5 unlocked it. Impossible to big business in three months; six months total from early to serious viability. Speed in that domain is a liability. For the human in the loop you want the frontier model; being 20% dumber than the next person is a problem. Routine operations can be cheaper open-source. A senior fintech operator he spoke to still used Claude Code or Codex for the human loop and cheaper models for PR review — logical, but the takeaway for him was still: augment yourself with the best you can afford.

Large firms pay down this debt over years. One strategy is to be a tech company: acceleration as core forever, Uber versus a taxi company, Amazon versus brick-and-mortar, enormous org cost, worth it. The other says acceleration is situational and will not rebuild around it. Those companies need partners. That is why Realfast exists. Thin products die when the model internalizes them. Build around invariants or you are buying time. His advice to founders is uncomfortably circular: do what I’m doing. He does not see fifty options. He sees people. America sucks high-quality humans in the loop into labs and big tech like oxygen. China will not serve the world. The other large pool is India, especially with the pipeline home that used to leave. The LeetCode-to-FAANG cannon is dead in the water. What is coming back is 1990s consulting: a serious hacker on a gnarly business problem. The talent has horsepower and the wrong cultural conditioning. Consulting requires social engineering — stakeholder alignment, in corpo. If you cannot do that, you have lost alpha.

The product/services line is not blurred. Durable partnerships still end in acquisition when you become core to someone else’s alpha. Structurally, he said, all product companies are a business team with a services backend bolted on. Nobody likes to admit it; the last logical step would be internal hourly billing to force talent allocation, and almost nobody takes it. At ThoughtWorks in 2005, a leftover of that hacker culture, the reasonable expectation was: new vertical, new stack, Java to .NET to Rails in 18 months, production, debt, speak the domain in four weeks. Very few people can do that now. Onboarding onto a wild ride every six to eight months is a different kick from ownership. US banking, he said, still has three old people holding the system together; when they die it is toast. Tools unlock those gnarly problems. The execution motion is unsolved.

You handed them a terrible app

Consumer companies live on a law: raise cognitive load, conversion dies. Introduce AI into a workflow and you increase the rate of human decisions. Most AI deployments therefore have terrible UX. Give a powerful agent to a normal user and you have given them something 100× worse than a shitty enterprise app. A tiny, scattered fraction of the workforce gets energized — don’t start Codex at 11 p.m. or you’ll be up until 3. Everyone else gets the anxiety of logging into a horrendous internal tool. ChatGPT works as a consumer app up through summarization and mild generation. Cross into “make me smarter” and engagement collapses. Top-down board pressure will close the first sale. Retention will be atrocious.

Realfast is not attacking giant companies yet. They operate roughly $50–500 million revenue, board-accessible, only projects the board cares about, because change management is the hard problem — crack it and you have a trillion-dollar company, not a billion. The idea that AI budgets have shifted from CIOs to operational managers who each own a metric, he called a myth: in B2B, outside support and some inside sales, nobody owns a consumer-style rollup; typical enterprises do not even know their metrics. The CIO 1% rollout, measure impact, then unlock 20/50/100%, is stuck. They cannot quantify impact to the CFO. Hence 99% of pilots look bad. Enterprise impact used to sit on multi-year Salesforce cycles with no iteration and a baked-in assumption that engagement would be a problem. Shelfware is a term he learned switching into this world. In consumer, an underutilized app is a lost customer. Boards now want 1% impact this quarter because a peer got AI-pilled at drinks. Either you are stuck in a partial rollout or you have 100% rollout, massive burn, and still cannot show impact. In big tech a leaderboard appears, token-maxing happens, quality control lands, it connects to performance reviews. In the enterprise, “show me impact” means the next quarterly checkpoint has to change, and almost no vendor is in the room talking about how you do PF.

Productivity, he said, is irrelevant. Operations research: local maxima, global minima. Speed one workstation and throughput can fall. Where companies understand this, you go from decision to production chip in nine months — what OpenAI has done. The depth of Claude Code in a year is humanly impossible from a cold start; Apple and Google’s annual OS cadence took years to install. Acceleration is evident in frontier labs and pockets of top-tier big tech. For everyone else, nothing. Not because multi-billion-dollar operators are clueless. Because they were not built from zero around acceleration as alpha. A 30-year-old $5 billion company is excellent fit to its environment. Change it and you cripple fitness. Innovator’s dilemma. Bolt on. He is the bolt-on.

The deepest market is BFSI, trillions, regulated for sane reasons. Realfast’s alpha is making the customer take their competition to the cleaners. Speed for its own sake is not the point. In that market acceleration is structurally gated. Shallower pools let you actually accelerate. He asked for one enterprise already destroying competitors with AI. Strongest revenue: support, inside sales, and, he added, AI girlfriends — which then tanked. Support is not better. It is cheaper, worse. Inside sales is 11 Labs calling about real estate. The inflection is when one enterprise in a vertical delivers crushing superiority — RFP in two days versus two weeks — and everyone else believes. Today demand is top-down. Belief is not in the bones.

They are not there yet as a vendor of that religion. A bizdev person built working CRM features on a Salesforce call in 18 minutes — not from scratch, on the platform — instead of going away to write a proposal. Scaled IT services lock 2–5 year contracts because software projects fail, risk sits with the customer, and short engagements go red. Boutiques can perform. Consumer-style workflow conversion metrics were never economically viable. They are now. Talent for that motion is different. Sell the crisp problem: March deadline, currently nine months late, can you do it in two months. Do not sell a leave-policy chatbot. The CFO will not call you mission-critical.

Wax on, wax off, with a robot lifting

IP has been taken without being paid for. Except at the high end, its value has collapsed. The art director in Delhi Belly who says make the banana smile 7% more: that craft has no economic value when you can tell the model. Websites, after years of no-code assault, are solved. Tacit knowledge in digital work is easier to instrument than in manufacturing. Onboarding a consultant used to take weeks. Eighty percent of tacit knowledge can sit in the repo, the harness, little scripts that used to be uneconomic to write. He believes hours, not days. The agent already writes commits in this team’s convention. Coordination roles thin out. Teaching a human augmented with an agent is an open question he is spending non-trivial time on. All his decade-and-a-half onboarding patterns are out. How do you wax on, wax off if the robot lifts the weights? Paying lakhs for an AI course is the old motion — honest, not a crypto scam — and the wrong mental model. Ask the AI to construct a pedagogy. He taught himself that way.

The economy assumed average talent was fine for average workflows. Arm good-enough people with agents and they lack the discipline to keep judging: twenty decisions a day that used to be three weeks. Fatigue, then slop, then compounding debt. Hamming them with an agent turns them into bad talent. The $30/hour Indian export band versus $5/hour freelance that fails even with micromanagement: the middle dropped to the bottom. Perfectly sensible national strategy — you cannot mint a million IIT graduates a year — suddenly invalidated. Welcome to the foothills of the singularity.

He is not pessimistic that India should have sovereign AI. He is pessimistic about competing with number one and number two. Two 15-trillion-plus economies, unique: Party, extreme free speech. At four to five trillion you are with France, Germany, Japan. The game is to be clean number three. Why does France have Mistral? Why isn’t Sarvam financed properly? Import substitution and blocking models (Fable) put you on a lower tier. Muscle compounds over decades; start now. Sharma-ji-ka-beta does not stay to make frontline math breakthroughs. Talent that will get you to number three is still in middle school. India sits in a trough: not China’s engineer-technocrat top-down, not America’s rugged individualism. Lawyers do not innovate; Western innovation comes from rebels, Chinese from technocrats. Academia here is teachers, not an arena. Exceptions: IISc, IIT Madras, Ather, Sarvam. Outside those zones, weak.

Bangalore as a city he will not forecast. Atoms businesses should get operational efficiency; how fast is unknown. Knowledge work is the backbone. Startups will not replace more than 2–3% of IT-services headcount. One-trick pony, not Bombay’s six industries. Technology is at the heart of the city’s culture; that argument ended in 1999, not 2026. IT services have trust, revenue, decades of Fortune 500 impact, and a fully vertically integrated talent supply chain. To become AI-native you cannot sprinkle a team. You 4x comp bands, drop review cycles to a quarter, stand up a structure that reports to the board, cannibalize without spooking jittery Indian investors. Brutal. He does not know how they navigate it.

The 180 he will own: for two and a half years Realfast’s mission was supercharging India’s median $30/hour export. Off the table. Either emergent $20 lovable prototypes with no reliability bar, or 99th-percentile talent you fight Amazon for. Personally, the subtler flip: an agent is not something you delegate to. It is you. It magnifies subconscious, not the story you tell about your conscious intent. Using it is deciding to move your hand. System two decides. System one executes. He pulled a 180 from anthropomorphizing to that.

Building observability in a digital world is trivial.

He asked to come back in six months. The operating horizon he named was three.