This is a submission for the OpenClaw Writing Challenge
everyone is saying openclaw would build my startup while i slept.
instead, i spent two weeks watching it burn through my api credits
while it asked the same question eight times in a row.
it wasn't thinking hard.
it was stuck in a loop, and i was the one paying $0.03 per token to watch it spin.
if you're currently babysitting your agent, watching it loop on simple tasks, or wondering if you should just go back to coding manually — i was there.
i almost gave up.
now i have openclaw running my morning briefings and handling database chores without me touching it.
the difference wasn't buying a better api tier. it was fixing these specific, stupid mistakes.
stop using your expensive model for everything
i had gpt-5.4 set as the default for every single task. heartbeat checks, file scans, cron jobs — all of it. i was asking a formula 1 car to deliver groceries.
openclaw lets you set up tiered model configs. here's what i switched to:
| task type | model |
|---|---|
| file reads, syntax checks, existence queries | haiku |
| actual coding tasks | sonnet |
| complex debugging, things that broke twice | opus |
my daily token spend dropped from 40,000 to around 1,500.
you can switch models mid-session with
/modelif you need to escalate, but most of the time, you don't.
your agent needs rules written in stone
out of the box, openclaw will:
- loop forever
- forget what it was doing
- rewrite your database schema because it misread a comment
you have to parent this thing with explicit, paranoid instructions.
i keep a workspace/skills/ folder full of SKILL.md files. these aren't suggestions. they're laws.
workspace/
├── skills/
│ ├── anti-loop.md
│ ├── USER.md
│ └── AGENTS.md
one file is literally called anti-loop.md and says:
"if you see the same error twice, stop and ask me. do not try a third variation."
another forces the agent to check USER.md before asking questions.
every assumption the agent makes is a potential landmine. openclaw doesn't know your database schema. it doesn't remember that you told it yesterday to never touch the auth module. write it down.
the agents that actually work are the ones with heavy custom instruction sets.
closing the chat kills the session
i told openclaw to optimize some queries and message me when done. closed my laptop. came back the next morning to find it had done nothing.
sessions die when you close the chat. they're stateful only while the window is open. when you reopen, you might get a summary, but the context, the stack, the "where was i" — gone.
what to do instead:
- use openclaw's cron jobs with isolated session targets
- these spin up fresh agent instances on a schedule, do one task, message you results, and die
- for one-off tasks, pair a simple
sqlitequeue with a cron that checks it hourly
# example cron entry
0 * * * * openclaw run --session-target=daily-briefing
don't try to maintain long-running thinking sessions. they break, they cost money, and they hallucinate when context gets long.
one working workflow beats five broken ones
i tried setting up email, calendar, telegram, web scraping, and reporting all at once. everything broke, and i couldn't tell which integration was failing.
start with one thing that hurts slightly every day.
i started with a morning briefing cron that:
- reads my calendar
- summarizes slack mentions
- messages me results
i got that working end-to-end — running without me touching it, messaging me reliably, failing loudly instead of silently — before i added anything else.
every new integration is a new failure mode. if things feel broken, run:
openclaw doctor --fix
half the "my agent is stupid" complaints are actually "my config is borked" problems.
compaction eats your memories
openclaw has a context window. when it fills up, the system compacts older messages — which means it forgets stuff.
i spent twenty minutes explaining my database schema once, then the agent compacted and hallucinated a new one. almost dropped a production table.
now i persist everything important:
| what | how |
|---|---|
| long-running task state |
JSON or YAML state files |
| user context | USER.md |
| behavior rules | AGENTS.md |
| architectural decisions | decision logs read at session start |
the less openclaw has to re-learn, the less it hallucinates.
chat quality and agent quality are different animals
i was using a model that gave beautiful, articulate chat responses. great reasoning. but it couldn't call tools to save its life. it generated malformed json and hallucinated function names.
models that actually work for agentic coding:
-
claude sonnet/claude opus gpt-5.4-
kimi k2.5via api
models to avoid:
-
deepseek reasoner— amazing at thinking, terrible at doing. it reasons beautifully about why your code is broken while generating completely broken tool calls. -
gpt-5.3 mini— cheap, but it skips steps and ignores tool results. multiple people have called it useless for agent work.
quick sanity test:
i) give your model three sequential tool calls.
ii) if it can't handle that without hand-holding,
iii) don't use it for autonomous work.
you're not bad at this. it's just early.
the gap between usual demos and daily use is real. when someone posts "my agent built a saas overnight," you're seeing the highlight reel. you're not seeing the three weeks they spent tuning prompts and debugging why openclaw kept trying to pay aws with monopoly money.
this stuff is genuinely hard right now. not "you need a cs degree" hard. just "the tools are immature" hard.
the people making it work are treating these agents like orchestras, not autopilots.
the four rules that changed everything for me:
- start with one cron job
- write one guardrail file
- use the cheap model
- save your state
it gets easier. don't give up before the compound interest kicks in.
Top comments (0)