Sitemap
3 min readSep 13, 2026

--

๐—”๐—œ ๐—–๐—ผ๐—ฑ๐—ถ๐—ป๐—ด ๐—ถ๐—ป ๐—š๐—ผ ๐—œ๐˜€ ๐—ก๐—ผ๐˜ ๐˜๐—ต๐—ฒ ๐—ฃ๐—ฟ๐—ผ๐—ฏ๐—น๐—ฒ๐—บ.
๐—จ๐—ป๐˜ƒ๐—ฒ๐—ฟ๐—ถ๐—ณ๐—ถ๐—ฒ๐—ฑ ๐—”๐—œ ๐—–๐—ผ๐—ฑ๐—ถ๐—ป๐—ด ๐—œ๐˜€.
I recently read Erwin Hermantoโ€™s article:
โ€œUsing AI Code Assistants for Go: What Works and What Still Failsโ€
Itโ€™s a good, pragmatic assessment of where Copilot, Cursor and Claude still struggleโ€Šโ€”โ€Šespecially with:
โ†’ concurrency
โ†’ context propagation
โ†’ hallucinated APIs
โ†’ code that looks correct, compilesโ€ฆand still contains bugs
I agree with the problems.
๐—•๐˜‚๐˜ ๐—œ ๐—ฑ๐—ผ๐—ปโ€™๐˜ ๐˜๐—ต๐—ถ๐—ป๐—ธ ๐˜๐—ต๐—ฒ ๐—ฟ๐—ถ๐—ด๐—ต๐˜ ๐—ฐ๐—ผ๐—ป๐—ฐ๐—น๐˜‚๐˜€๐—ถ๐—ผ๐—ป ๐—ถ๐˜€ โ€œ๐—”๐—œ ๐—ถ๐˜€๐—ปโ€™๐˜ ๐—ฟ๐—ฒ๐—ฎ๐—ฑ๐˜† ๐—ณ๐—ผ๐—ฟ ๐˜๐—ต๐—ฒ ๐—ต๐—ฎ๐—ฟ๐—ฑ ๐—ฝ๐—ฎ๐—ฟ๐˜๐˜€.โ€
The bigger issue is that we are still evaluating AI as if it were autocomplete.
The question should no longer be:
โŒ โ€œCan an LLM generate perfect production Go code from one prompt?โ€
It canโ€™t. Not reliably.
The better question is:
โœ… โ€œCan an agent understand the repository, implement a change, test it, detect problems, fix them, challenge its own implementation and only then create a PR?โ€
That is a completely different engineering model.
โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”
๐—ง๐—ฎ๐—ธ๐—ฒ ๐—ฐ๐—ผ๐—ป๐—ฐ๐˜‚๐—ฟ๐—ฟ๐—ฒ๐—ป๐—ฐ๐˜†.
If an AI creates a data race, the solution shouldnโ€™t be:
โ€œHopefully a senior developer notices it during review.โ€
Run:
go test -raceย ./...
Add static analysis.
Add adversarial tests.
Let a dedicated reviewer agent inspect concurrency.
Then make those checks mandatory in CI.
๐—ง๐—ต๐—ฒ ๐—บ๐—ผ๐—ฑ๐—ฒ๐—น ๐—ฑ๐—ผ๐—ฒ๐˜€๐—ปโ€™๐˜ ๐—ต๐—ฎ๐˜ƒ๐—ฒ ๐˜๐—ผ ๐—ฏ๐—ฒ ๐—ฝ๐—ฒ๐—ฟ๐—ณ๐—ฒ๐—ฐ๐˜.
The system around it has to make failures difficult to survive.
โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”
๐—–๐—ผ๐—ป๐˜๐—ฒ๐˜…๐˜ ๐—ฝ๐—ฟ๐—ผ๐—ฝ๐—ฎ๐—ด๐—ฎ๐˜๐—ถ๐—ผ๐—ป?
Donโ€™t hope Claude remembers the correct pattern.
Make it a project rule:
โ€ข propagate the callerโ€™s context
โ€ข respect cancellation and deadlines
โ€ข donโ€™t introduce context.Background() inside request flows
โ€ข derive child contexts correctly
Put those rules into CLAUDE.md / project rules.
Then enforce them through static analysis and tests.
That turns โ€œprompting adviceโ€ into an ๐—ฒ๐—ป๐—ด๐—ถ๐—ป๐—ฒ๐—ฒ๐—ฟ๐—ถ๐—ป๐—ด ๐—ฐ๐—ผ๐—ป๐˜๐—ฟ๐—ผ๐—น.
โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”
๐—›๐—ฎ๐—น๐—น๐˜‚๐—ฐ๐—ถ๐—ป๐—ฎ๐˜๐—ฒ๐—ฑ ๐—”๐—ฃ๐—œ๐˜€?
The agent should not rely on what it โ€œremembersโ€.
It should inspect:
โ†’ the repository
โ†’ go.mod
โ†’ package source
โ†’ actual documentation
โ†’ existing usage patterns
Then verify:
go buildย ./...
go vetย ./...
staticcheckย ./...
go testย ./...
go test -raceย ./...
govulncheckย ./...
If verification fails, the task isnโ€™t finished.
The agent fixes the problem and runs the checks again.
๐—š๐—ฒ๐—ป๐—ฒ๐—ฟ๐—ฎ๐˜๐—ฒ๐—ฑ ๐—ฐ๐—ผ๐—ฑ๐—ฒ ๐—ถ๐˜€ ๐—ป๐—ผ๐˜ ๐˜๐—ต๐—ฒ ๐—ฑ๐—ฒ๐—น๐—ถ๐˜ƒ๐—ฒ๐—ฟ๐—ฎ๐—ฏ๐—น๐—ฒ.
๐—ฉ๐—ฒ๐—ฟ๐—ถ๐—ณ๐—ถ๐—ฒ๐—ฑ ๐—ฏ๐—ฒ๐—ต๐—ฎ๐˜ƒ๐—ถ๐—ผ๐—ฟ ๐—ถ๐˜€.
โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”
And there is another important change:
๐—ข๐—ป๐—ฒ ๐—”๐—œ ๐—ฑ๐—ผ๐—ฒ๐˜€๐—ปโ€™๐˜ ๐—ต๐—ฎ๐˜ƒ๐—ฒ ๐˜๐—ผ ๐—ฑ๐—ผ ๐—ฒ๐˜ƒ๐—ฒ๐—ฟ๐˜†๐˜๐—ต๐—ถ๐—ป๐—ด.
A more mature Agentic Engineering workflow can look like this:
๐—ฅ๐—ฒ๐—พ๐˜‚๐—ถ๐—ฟ๐—ฒ๐—บ๐—ฒ๐—ป๐˜๐˜€
โ†“
๐—œ๐—บ๐—ฝ๐—น๐—ฒ๐—บ๐—ฒ๐—ป๐˜๐—ฎ๐˜๐—ถ๐—ผ๐—ป ๐—”๐—ด๐—ฒ๐—ป๐˜
โ†“
๐—ง๐—ฒ๐˜€๐˜ ๐—”๐—ด๐—ฒ๐—ป๐˜
โ†“
๐—–๐—ผ๐—ป๐—ฐ๐˜‚๐—ฟ๐—ฟ๐—ฒ๐—ป๐—ฐ๐˜† / ๐—–๐—ผ๐—ฟ๐—ฟ๐—ฒ๐—ฐ๐˜๐—ป๐—ฒ๐˜€๐˜€ ๐—ฅ๐—ฒ๐˜ƒ๐—ถ๐—ฒ๐˜„
โ†“
๐—ฆ๐—ฒ๐—ฐ๐˜‚๐—ฟ๐—ถ๐˜๐˜† ๐—ฅ๐—ฒ๐˜ƒ๐—ถ๐—ฒ๐˜„
โ†“
๐——๐—ฒ๐˜๐—ฒ๐—ฟ๐—บ๐—ถ๐—ป๐—ถ๐˜€๐˜๐—ถ๐—ฐ ๐—–๐—œ
โ†“
๐—ฃ๐˜‚๐—น๐—น ๐—ฅ๐—ฒ๐—พ๐˜‚๐—ฒ๐˜€๐˜
โ†“
๐—›๐˜‚๐—บ๐—ฎ๐—ป ๐—๐˜‚๐—ฑ๐—ด๐—บ๐—ฒ๐—ป๐˜
None of those agents are infallible.
Thatโ€™s exactly the point.
We should not design a development process where correctness depends on one model being right.
โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”
Thereโ€™s also an important feedback loop:
If a human reviewer discovers:
โ€œEvery goroutine needs a clear termination condition.โ€
or
โ€œNever create context.Background() inside HTTP handlers.โ€
or
โ€œConcurrent processing must preserve ordering unless explicitly specified otherwise.โ€
โ€ฆthat knowledge should not die inside a PR comment.
๐—ฃ๐˜‚๐˜ ๐—ถ๐˜ ๐—ถ๐—ป๐˜๐—ผ ๐˜๐—ต๐—ฒ ๐—ฝ๐—ฟ๐—ผ๐—ท๐—ฒ๐—ฐ๐˜โ€™๐˜€ ๐—”๐—œ ๐—ฟ๐˜‚๐—น๐—ฒ๐˜€.
Now the next task starts with that knowledge.
That is how the engineering system improves over time.
โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”
The original article is right about one very important thing:
The hard parts of software engineering do not disappear just because an LLM can generate the code. ๏ฟฝ
Using AI Code Assistants for Go_ What Works and What Still Fails _ by Erwin Hermanto _ Sep, 2026 _ Medium.pdf
But AI changes ๐˜„๐—ต๐—ฒ๐—ฟ๐—ฒ the engineering effort goes.
Less time manually typing implementation.
More time defining:
โ€ข requirements
โ€ข architecture
โ€ข invariants
โ€ข acceptance criteria
โ€ข tests
โ€ข verification
โ€ข security boundaries
โ€ข automated quality gates
โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”โ”
The future is not:
โŒ AI writes code โ†’ humans trust it
And I donโ€™t believe the long-term answer is:
โŒ AI writes code โ†’ humans manually inspect every generated line forever
The direction is:
๐—”๐—ด๐—ฒ๐—ป๐˜๐˜€ ๐—ถ๐—บ๐—ฝ๐—น๐—ฒ๐—บ๐—ฒ๐—ป๐˜.
๐—”๐—ด๐—ฒ๐—ป๐˜๐˜€ ๐—ฐ๐—ต๐—ฎ๐—น๐—น๐—ฒ๐—ป๐—ด๐—ฒ.
๐—ง๐—ผ๐—ผ๐—น๐˜€ ๐˜ƒ๐—ฒ๐—ฟ๐—ถ๐—ณ๐˜†.
๐—–๐—œ ๐—ฒ๐—ป๐—ณ๐—ผ๐—ฟ๐—ฐ๐—ฒ๐˜€.
๐—›๐˜‚๐—บ๐—ฎ๐—ป๐˜€ ๐—ท๐˜‚๐—ฑ๐—ด๐—ฒ.
The goal is not to make AI so trustworthy that verification becomes unnecessary.
๐—ง๐—ต๐—ฒ ๐—ด๐—ผ๐—ฎ๐—น ๐—ถ๐˜€ ๐˜๐—ผ ๐—ฏ๐˜‚๐—ถ๐—น๐—ฑ ๐—ฎ๐—ป ๐—ฒ๐—ป๐—ด๐—ถ๐—ป๐—ฒ๐—ฒ๐—ฟ๐—ถ๐—ป๐—ด ๐˜€๐˜†๐˜€๐˜๐—ฒ๐—บ ๐˜„๐—ต๐—ฒ๐—ฟ๐—ฒ ๐˜๐—ฟ๐˜‚๐˜€๐˜ ๐—ถ๐˜€ ๐—ป๐—ผ ๐—น๐—ผ๐—ป๐—ด๐—ฒ๐—ฟ ๐˜๐—ต๐—ฒ ๐—ฝ๐—ฟ๐—ถ๐—บ๐—ฎ๐—ฟ๐˜† ๐—บ๐—ฒ๐—ฐ๐—ต๐—ฎ๐—ป๐—ถ๐˜€๐—บ ๐—ณ๐—ผ๐—ฟ ๐—ฎ๐—ฐ๐—ต๐—ถ๐—ฒ๐˜ƒ๐—ถ๐—ป๐—ด ๐—ฐ๐—ผ๐—ฟ๐—ฟ๐—ฒ๐—ฐ๐˜๐—ป๐—ฒ๐˜€๐˜€.
And that principle applies to human-written code too.
โ€”
Inspired by / response to:
Erwin Hermanto
โ€œUsing AI Code Assistants for Go: What Works and What Still Failsโ€
#AgenticAI #SoftwareEngineering #Golang #ClaudeCode #AICoding #AIEngineering

--

--

Golang +
Golang +

Written by Golang +

#golang news, interesting stuff, blog links and golang jobs.