The notebook
Two kinds of writing. Bugs that produced no error message and cost me real time in a real codebase — and first-person accounts of learning to ship software without ever having been a developer. All of it first-hand.
MIGI is a fleet of AI agents built by Suman Debnath, running since 8 July 2026 at forty to fifty agent runs a day. It has cost under five dollars in total, against published estimates of $185 to $480 a month for a comparable personal stack, because the paid model is a fallback rather than the default.
Read the article
Working notes, newest first.
Worth your time
An ancestor with overflow: hidden becomes the sticky element's scroll container. That container never scrolls, so the sticky child has nothing to stick to and scrolls away with the page instead. Nothing errors and no warning appears. Check every ancestor for overflow-hidden before touching the sticky element's own CSS.
Also worth it
Three rules do most of the work: never let a key appear in code, keep environment files out of version control, and hold production values in the host's environment settings. Rotation is the part most people skip. Push protection catches known key formats only, so it is a net rather than a guarantee.
The notebook's shape, rather than its order.
The rest of the notebook, newest first.
The useful test is not whether a product calls somebody else's model, because most do, including good ones. It is what would remain if the model were removed. A product with its own evaluation, defined behaviour when the answer is wrong, real integration and data you can export is a product. One with none of those is an interface charging a markup on an API.
Between 21 June and 1 July 2026, Suman Debnath created eight repositories. Six of them stopped receiving commits within three days; the two that did not became an agent fleet still in active development two months later. The mistake was not speed — it was treating a working first version as a finished product.
Less syntax than expected and more operations than anyone warns you about. The gap is not writing code — an assistant does that — it is knowing where data lives, how deployment differs from a machine that works, what happens when something breaks in front of a user, and being able to read a change well enough to accept or reject it.
Four questions decide it: can an agent find the site, fetch it, read it without executing JavaScript, and recover when it hits a URL that does not exist. Most sites fail on the third for a dull reason — the content or the structured data only exists after the client bundle runs, and many fetchers never run it.
AI-native describes someone whose default working method is building with these systems, rather than someone who added them to an existing method. The claim is checkable: ask what they shipped, what they did when the model was wrong, what they refused to automate, and how they verify output. Tool familiarity answers none of those.
Start with the highest-frequency, lowest-judgement work — reporting, data gathering, asset variants and routine checks — because that is where hours accumulate without any decision being made. The selection rule that matters most is error visibility: automate where a mistake is obvious and cheap, and never where it would be silent and expensive.
Search optimisation competes to be ranked among ten results; answer optimisation competes to be the one passage quoted. The unit shifts from the page to the self-contained block, the currency shifts from keywords to resolved entities, and the outcome becomes largely unmeasurable — there is no rank to track and the same question returns different answers on different runs.
A weekend produces a working version, not a finished product — those are separated by weeks of unglamorous work. What makes the weekend possible is severe scope discipline and research before the first prompt. What makes it defensible is refusing to cut the handful of things that are cheap now and expensive later.
Taste is the ability to choose between options that are all technically correct, and it becomes the scarce skill once generation is cheap. A model optimises whichever metric it is given. Deciding which metric is the right one — and recognising when a measurable win is the wrong trade — is the part that cannot be delegated.
Probably not for search visibility. Google has publicly declined to support llms.txt, and reported monitoring of AI crawler traffic through 2026 found direct fetches to be negligible. Where it does get read is agent-side — coding assistants and desktop agents retrieving a domain because a user named it. Worth having if it is generated, not worth maintaining by hand.
Named tools are not skills, because everyone has access to the same tools and the claim cannot be checked. What survives scrutiny is anything with an artefact or a measured outcome attached: a process automated with the time saved stated, a system built and linked, or a described method for judging whether an output is correct.
React StrictMode mounts every component, runs cleanup, then remounts. A module-level `if (initialised) return` guard is set on the first mount, survives the cleanup because it lives outside the component, and blocks the second mount from ever re-attaching its listeners. The feature is then dead for the whole dev session while production works fine.
Open to meaningful collaborations, AI-native systems, product strategy, and future-focused conversations.
“Human instinct. AI amplification.
Systemic execution.”
Suman Debnath
·Brand Marketing Leader & AI Product Builder
© 2026
This site records visit data — pages viewed, time and scroll depth, device, your IP address and the approximate location and network provider derived from it — and sends it to me privately. It also runs Google Analytics and Vercel Analytics. Full detail and how to opt out.
This site, its code and its content are © 2026 Suman Debnath. All rights reserved — none of it is open source, and copying it needs permission first. Terms of use.
