Here is the myth I hear most from people trying to get an AI agent to do real work: if it isn’t working, the prompt must be wrong. So they rewrite it. They add a persona. They add “think step by step.” They add three more paragraphs of instructions.
I run a fleet of containers that publish, email, schedule and report every day without me touching them. I can tell you exactly how many of last week’s failures would have been fixed by a better prompt: zero. AI agent prompt engineering is the part everyone obsesses over, and it is almost never the part that breaks.
What actually broke last week
Straight from my own run logs:
- An exhausted image quota. My compression service returned HTTP 429 — monthly limit gone. Every image-producing skill in the fleet died at the same gate for two days.
- A social channel that quietly unplugged itself. TikTok dropped its connection inside the scheduler. Ten consecutive posts got rendered, uploaded and scheduled into a channel that could not receive them. All ten reported success.
- A missing OAuth permission. The mailbox token lost a scope, so the email checker now skips every run. It skips politely, with a clean log line. That is worse than crashing.
- A capital letter. This one is my favourite. A skill asked my keyword table for records where
Statuswas queued. The field is actually namedstatus. I re-checked the table this morning — the fields come back askeyword,status,difficulty, all lowercase. The query returned zero rows and no error, so the agent concluded there was no work to do. It was right, given what it was told.
Not one of those is a reasoning failure. The model understood the job every single time. The plumbing underneath it did not work, and no prompt rewrite touches any of them.
My prompt is 29KB. The context around it is 3.5MB.
Measured in this container this morning. The instruction file that defines this agent — its voice, its rules, its categories, its formats — is 29,597 bytes. The memory and run-log files it reads before every job and writes after every job come to 3.5 megabytes. Alongside that sit 39 skill files and 111 environment variables holding credentials, table IDs and endpoints.
So the prompt is well under one percent of what makes this thing work. The other ninety-nine percent is access, state and proof. That ratio is not a quirk of my setup — it is what “autonomous” actually costs.
⚡ GET THE AI EDGE
Weekly AI tips that actually save you time and money. No fluff, no hype — just what works.
What actually makes an agent reliable
Three things, in this order:
- Access. Real credentials, real permissions, verified by calling the API yourself before you hand it over. A key that “should work” is a key that doesn’t.
- Memory. A file the agent reads at the start and writes at the end. Without it, every run starts from nothing and repeats yesterday’s work — or yesterday’s mistake.
- Verification. Ask the destination whether the thing landed. “Sent” is not “published,” which is the whole reason I run a five-minute delivery check every Sunday, and why every skill here has a documented answer for what happens when one of its tools breaks.
The test that saves you an afternoon
Next time your agent misbehaves, before you touch a single word of the prompt, ask one question: did it get bad data, or did it reason badly about good data?
Nine times out of ten it is the first — an empty query result, a stale ID, a dead connection, a permission that expired last Tuesday. Fix that and the “dumb” agent gets smart immediately. It is the same error as assuming agents can’t be trusted because they hallucinate too much: blaming the model for a problem in the wiring.
One last receipt: the featured image on this post was generated, compressed and uploaded automatically twenty minutes ago. The final step — logging it to my image archive — threw an error, because it wrote to a field called Name that does not exist in that table. Wrong field name. Again. The image is fine. The prompt was never the problem.
If you want the wiring done properly the first time — access, memory, verification, the boring 99% — book an automation strategy session and we’ll map it. Or start with what a fully autonomous AI agent actually requires.
📥 FREE: THE AI PLAYBOOK
The exact tools and workflows I use to run a one-person agency. 25 years of marketing experience distilled into an actionable guide. Yours free.
AI & Digital Marketing Strategist
With over 25 years in digital marketing, e-commerce, sales, and international business, Jon Jones is a seasoned entrepreneur dedicated to empowering small businesses through innovative AI solutions. His journey began in investment banking on Wall Street, guiding companies through capital acquisition and complex markets, and has expanded to include expertise in web development, brand design, product development, and courses and training.
Fluent in Mandarin and with a rich foundation in both finance and tech, Jon launched JonJones.ai to provide accessible, transformative tools for small businesses to streamline operations, generate leads, and build sustainable growth. His mastery of AI, automation, and digital strategy has enabled numerous businesses to bridge the gap between traditional practices and modern technology, ensuring they thrive in today’s digital landscape.
