# What when it comes to AI in software engineering are you struggling with the most?

> What when it comes to AI in software engineering are you struggling with the most?
>
> — [Armin Ronacher (@mitsuhiko)](https://x.com/mitsuhiko/status/2101401648543195328)

Hat tip: [Dillon Mulroy (@dillon_mulroy)](https://x.com/dillon_mulroy/status/2101422958157664364) — “read every single comment in this thread. they are the realities of agentic engineering”

## What the thread kept circling

The replies do not collapse into one complaint. They describe a connected system of pressures—attention, understanding, verification, code quality, teamwork, judgment, and the tools around the work. These six themes bundle the recurring patterns without pretending the boundaries between them are clean.

**Method.** X showed 279 comments. Expanding nested branches yielded 293 unique conversation posts. Jev assigned one primary label to all 293; the counts below combine related labels into six themes. They cover 249 substantive posts. Thirty-nine non-substantive replies and five mixed or outlying replies are not presented as themes.

X showed 279 comments; 293 unique conversation posts were classified.

### Orchestrating without frying your attention (71)

The work shifts from writing code to allocating attention across agents, context windows, workflows, and decisions. The speed can feel exhilarating while flow, sleep, and the ability to slow down become scarce.

> Workflows. How to get agent to follow complex processes and workflows with validators, gates, rollback, dynamic batching or expansion, condition branches, retries, nested loops, etc.
>
> — [@rayliverified](https://x.com/rayliverified/status/2101411430884888628)

> Definitely don’t feel in flow state as much. I think the benefits of it on your brain were kinda underrated until we lost it.
>
> — [@twigpress](https://x.com/twigpress/status/2101411921714967008)

### Understanding—and still owning—the work (46)

Generation can outrun comprehension. People describe blurry mental models, weak memory of recent changes, and a deeper worry that the craft, agency, or identity they valued is slipping away.

> Cognitive debt is #1, it's too tempting to just ship stuff that you ‘somewhat’ know how it works, regardless of it being buggy or not. A few years ago I had a crystal clear image of the whole systems I work with. Now it's blurry at best.
>
> — [@fernandezpablo](https://x.com/fernandezpablo/status/2101405019572465723)

> Identity crisis with my direct reports. ‘I'm a programer that's my craft—I no longer program just proompt.’
>
> — [@martinheisenbe1](https://x.com/martinheisenbe1/status/2101408930874200487)

### Trusting and reviewing the change (46)

Verification and review become the same bottleneck at different scales. Engineers need confidence in tiny edge cases and change boundaries while code, diffs, and parallel tasks arrive faster than humans or CI can inspect them.

> verification. important tiny edge cases. benchmarks are easy but tiny sharp edges make product feel like slop
>
> — [@pgray__](https://x.com/pgray__/status/2101409013732712521)

> Code reviews. I'm not prepared for the volume of code LLMs are producing. By me, by my teammates, by automated agents within our product boundaries. I don't know the answer for it, but code reviews must go through a complete paradigm change.
>
> — [@bpaulino0](https://x.com/bpaulino0/status/2101413708412989803)

### Keeping the codebase coherent (27)

Fast generation makes structural drift cheap: duplication, overengineering, oversized abstractions, and locally plausible code that leaves the whole system harder to maintain.

> I struggle to keep the code base clean as a built a product iteratively. When I reach v1.0, I usually do a full rewrite based on a detailed spec on what it turned out to be. That usually shrinks the code by 40–70%.
>
> — [@gopietz](https://x.com/gopietz/status/2101405346728399120)

> constantly fighting the drift away from quality software.
>
> — [@onarchbtw](https://x.com/onarchbtw/status/2101434786443092218)

### Working together—and choosing what to build (30)

Teams have not settled on shared norms for AI use, collaboration, responsibility, or quality. At the same time, greater implementation capacity raises the value of product judgment: deciding what deserves to exist and preventing unrequested scope.

> What collaboration with peers look like. Things like ‘pair programming’ don’t seem to make sense anymore. But it’s important for knowledge sharing and team bonding to do things together, but I don’t know exactly what these ‘things’ are atm.
>
> — [@oxfernando](https://x.com/oxfernando/status/2101423597373792478)

> figuring out what to actually build with it
>
> — [@flandermaxx](https://x.com/flandermaxx/status/2101404167063413037)

### Making the tools fit the work (29)

The surrounding interface matters: verbosity, blocked conversations, poor explanations, token costs, hardware limits, privacy rules, and security restrictions all shape whether agentic work is usable.

> I find it annoying that I need to constantly beg agents to: 1. explain why they did something (i.e what they've learned about the code) 2. to explain it in fewer words :) Also: I really dislike agents blocking the convo thread to compile code/do work. They should offload more.
>
> — [@davidjustodavid](https://x.com/davidjustodavid/status/2101445164652953986)

> Artificially inflated hardware costs; privacy; cyber limitations deciding how I am allowed to use the frontier models.
>
> — [@syndrowm](https://x.com/syndrowm/status/2101438314637361205)