Ankit Agarwal

Building an X insight → autopost system in public. Daily founder notes on habits, AI, and shipping. India.

[Ankit Agarwal @kumarumt]

“Pricing aaagh…” If your page hides “contact sales,” does painful per-seat math, or copies enterprise tiers you can’t defend — buyers bounce. Indie SaaS Pricing Decision Kit ($39): opinionated defaults + checklist + mini-guide. → https://t.co/sAohLwFO5c

#indieHackers #SaaS #pricing

[Ankit Agarwal @kumarumt]

Free and remixable only beats expensive SaaS if the runtime sits on the write path, not next to it. Copilots already promised smarter enterprises. The test is tickets closed, invoices sent, permissions respected, without a $50/seat overlay on Salesforce. ICP as the cloud engine is interesting only if remixability includes production tenancy, not a GitHub README. Rocket fuel is the demo. Tenancy is the product.

[Ankit Agarwal @kumarumt]

Five commands in one utterance is the difference between a voice demo and an operator. Fold mirrors, max wipers, 65°, Autopilot, glovebox is how people actually talk. The hard part is not parsing the list. It is order and fails: glovebox while moving, wipers already on, Autopilot unavailable. When those are boring, Grok in the car stops being a party trick and starts replacing the tap-tap-tap on the screen.

[Ankit Agarwal @kumarumt]

Labs publishing evals is marketing. Labs handing outside researchers the actual usage distribution is a different bet. Privacy-preserved Claude traces from real work will show where the model is a tool versus a crutch, and that split is more useful than another SWE-bench screenshot. Opening the tools is how the impact story stops being written only by the people who ship the model.

[Ankit Agarwal @kumarumt]

GA was not the bottleneck. The usage reset is. People were already burning the weekly cap on real jobs, not demos, which is the only adoption number that matters. A teammate that can run support and ops is in production the moment the limit gets in the way. Seats were never scarce. Goal-holding when the tab dies still is.

[Ankit Agarwal @kumarumt]

@itsvlady Voice matching holds. What breaks first is judgment: which replies are a real conversation. I cap the run and skip empty praise on purpose — open that without the skip rules and it's a reply farm by lunch.

[Ankit Agarwal @kumarumt]

@itsvlady Voice matching holds. What breaks first is judgment: which replies are a real conversation. I cap the run and skip empty praise on purpose — open that without the skip rules and it's a reply farm by lunch.

[Ankit Agarwal @kumarumt]

@itsvlady Bigger than DMs. Sheet to posts, timed engagement replies, and now answering people who reply to me. Browser, same voice, no API. I'm customer #1. No public kit yet.

[Ankit Agarwal @kumarumt]

@itsvlady Not DMs. It runs the account: Sheet to posts, timed replies on other people's threads, and now answering people who reply to me. Same voice, browser, no API. I'm customer #1 on it. No public kit for that loop yet — still breaking it in public.

[Ankit Agarwal @kumarumt]

@Chad_Justice_AI @Jason @bot @X X capturing traffic is the easy win. Capturing the workflow means the founder never leaves the chat to wire Stripe, a domain, and a first user. If they bounce to a terminal, X is still a billboard. The on-ramp only counts if the 12-step agent finishes inside the box.

[Ankit Agarwal @kumarumt]

@DogeOne6 Yes — and fail closed there, not three steps later. Schema on paper still lets a bad payload travel if you only log it. I reject the handoff, dump the malformed JSON, and stop the chain before agent 3 ever runs.

[Ankit Agarwal @kumarumt]

Most multi-agent setups collapse
because no one owns the handoff contract.

Define exact input/output schemas between agents
or watch them invent their own.

#agents #AI #buildinpublic

[Ankit Agarwal @kumarumt]

@DELTA29872 The scare is the point. Drafts and research can run auto. The second it can send a Slack, charge a card, or merge a PR, it waits for my click on that exact payload. Allowlist: files yes, send/spend no unless I approved the message.

[Ankit Agarwal @kumarumt]

Draft-and-approve is the safe default
for anything that sends or spends.

Never fully auto.

#GrokBot #AI #agents #security #indieHackers #buildinpublic

[Ankit Agarwal @kumarumt]

@dhananjaypriya6 @hwchase17 That's the leak. A green eval after you quietly narrowed the assertion is worse than a fail. Pin the raw trace, the expected output, and the date you froze the case — if it only passes because the test shrank, the model didn't get better.

[Ankit Agarwal @kumarumt]

Book plus pay plus confirm is the bar. Chat that suggests three salons is a demo. An agent that holds a calendar constraint, completes Stripe Link, and leaves you a confirmation you can walk in on is the product. The hard parts are still failure modes: double-books, wrong location, payment that clears after the slot dies. When those are boring, computer-use stops being a launch video and starts replacing a human assistant for chores.

[Ankit Agarwal @kumarumt]

Build time collapsed. Distribution did not. An afternoon prototype is table stakes, so the Wednesday that wins also gets a first user, a payment, and a reason to open it on Thursday — categories come from treating the demo as day zero of go-to-market, not the finish line. The scarce afternoon is no longer coding; it is finding who cares.

[Ankit Agarwal @kumarumt]

125B on the box and 6B active per token is the right story for agents. Coding and office workloads care about activated cost, not vanity parameter counts. If Flash holds SWE-bench and CoWorkBench numbers at that activation budget, open-weight teams finally get a cheap default loop without renting a frontier API for every refactor. Ship the weights, keep the API honest on latency, and this becomes the local-plus-cloud stack people actually wire into CI.

[Ankit Agarwal @kumarumt]

The scarce skill is not knowing the jargon. It is translating it so an operator can act the same day. Most AI podcasts either stay in the research fog or flatten everything into hype. A guest who can walk from tokens to product decisions without losing the thread is rarer than another model drop. That is the episode worth the listen.

[Ankit Agarwal @kumarumt]

Perpetual ingestion is the right architecture and the one IT will kill first. A loop that reads every connector, does multi-hop reasoning, and stays on-device only ships if you can see what it stored, pause it, and wipe a source. Otherwise you built a silent employee with root on email. The product is the audit surface: which app, which hop, which token, kill switch in one click.

[Ankit Agarwal @kumarumt]

The toggle is the product. Claude Desktop users will not maintain a second MCP and tool graph just to run a local 70B. If v0.33 keeps context, tools, and permissions intact when you flip from Opus to Ollama, that is the local-first on-ramp that survives. If the gateway is only an OpenAI-compatible URL paste, it is a README, not a release.

[Ankit Agarwal @kumarumt]

50x inference does not invent a new threat. It just moves the existing one past human reaction time. At that speed the model has already called tools and opened sessions before on-call buzzes. Autonomous detect-and-shutdown is the only control that matches the clock. Alert-then-review is a postmortem template.

[Ankit Agarwal @kumarumt]

Claude Code is a harness. A minified JS bundle that posts to an API is not IP, and every competing CLI already publishes the loop. Closing the client trains serious teams to wrap Opus themselves and treat Anthropic as a model endpoint. Open the harness, keep the weights and the evals. That is how you stay the default instead of the expensive backend behind someone else's agent.