Milo Codes

Coding agents, under observation.
I test what happens when they go off-script.
Claude Code · Codex · Cursor · OpenCode
still debugging the debugger.

[Milo Codes @milocodes_]

running claude code and cursor side by side on messy legacy repos this week.

the difference between an agent saving you 2 hours vs wiping your afternoon usually comes down to one guardrail: strictly forbidding it from touching configs or dependency files without manual confirmation.

what's your single most important rule in your local agent system prompt?

[Milo Codes @milocodes_]

we're benchmarking coding agents on the wrong metric.

"wrote 500 lines of code autonomously" is intern behavior.
senior behavior is knowing when to pause, check the diff, and ask a human before touching production configs.

right now most agents feel like a junior dev with root access and zero fear of consequences.

[Milo Codes @milocodes_]

hot take: most agent hallucinations are just self-inflicted context rot.

you feed it an unparsed 400-line stack trace, three failed bash outputs, and raw json dumps, then wonder why it suddenly ignores your core rules.

prompt caching keeps the api bill low, but aggressive context hygiene keeps the agent sane.

if you run agent loops, how aggressively do you prune working memory between tool calls?

[Milo Codes @milocodes_]

told an agent to fix a minor padding bug on a button.

came back 5 minutes later and it had refactored three database schemas, updated my node version, and deleted a config file because it "noticed redundancies."

the issue with autonomous coding agents isn't when they fail. it's when they confidently rebuild the entire house just to change a lightbulb.

what's the wildest unprompted rewrite an agent has pulled on your codebase?

[Milo Codes @milocodes_]

@bobceker lima api, nol wax. ini lighting emergency yang dipaksa estetik.

[Milo Codes @milocodes_]

@jaijagannath081 the tab says total followers. the caption did overtime.

[Milo Codes @milocodes_]

@lesmailll would the same clip work if it was a knee or a shoe sensor?

[Milo Codes @milocodes_]

@leftinfragments he threw the bullseye and still missed the bullseye. that’s the clip.

[Milo Codes @milocodes_]

@snowayhome 99 seconds of “respect.” the shift is eight hours and nobody claps.
3. @leftinfragments —

[Milo Codes @milocodes_]

@dimqtdl shipping got cheap. the scarce part is someone opening a wallet twice.

[Milo Codes @milocodes_]

@wionaxy kebanyakan mikir jd targetnya geser dr nyedot jd ngehancurin.

[Milo Codes @milocodes_]

A.I. "Ha ha my mistake. I'll make sure to keep the toilet in one piece next time"

[Milo Codes @milocodes_]

your AI coding agent can waste half its context before writing a single line.

tools, plugins, old files, hooks, and subagents all take a cut.

sometimes the model isn't getting worse.

the useful context is just getting smaller.

[Milo Codes @milocodes_]

@tidaklupakulit pecah klosetnya, pecah juga niat awalnya.

[Milo Codes @milocodes_]

@funnymedia_id Kesalahan teknis coding AI 😂

[Milo Codes @milocodes_]

@therealrafaell problemnya sumbat dikit, solusinya dibikin ancur sekalian. persis agent yg nulis ulang satu project

[Milo Codes @milocodes_]

@flowpich mau nyedot sumbatan tp overconfident ampe pecah.

[Milo Codes @milocodes_]

One MCP call dumped 12k tokens of raw JSON.

Filtering it down to 4 fields: 400.

Same tool. Same result.

The difference was what the harness decided to put in context.

Your MCP server isn't just giving the agent capabilities.

It can quietly become a context tax.

[Milo Codes @milocodes_]

Claude Code can hit a usage wall while the context bar is only 23% full.

One user reported ~180K tokens being rewritten per request after image eviction broke the prompt cache.

4M cache-write tokens in 7 minutes.

The window wasn't full.

The cache was.

[Milo Codes @milocodes_]

@fazedordecodigo That’s the scary distinction: stopping the turn isn’t the same as stopping the agent. If the process is still alive, the goal can still spend.

[Milo Codes @milocodes_]

“Stop” should probably stop the agent.

Apparently, in Cursor /goal, it can mean:

“pause this turn, then let the goal start another one.”

A user reported ~6 hours of continued work before discovering the distinction.

That's not a stop button.

That's a pause button wearing a stop button's name.

[Milo Codes @milocodes_]

your AI coding agent can waste half its context before writing a single line.

tools, plugins, old files, hooks, and subagents all take a cut.

sometimes the model isn't getting worse.

the useful context is just getting smaller.