AI agents
Prompting Is Not a Strategy
Everyone who spent 2024 collecting prompts has quietly stopped. Not because prompts stopped working, because a prompt was never the unit that produced value.
Key takeaways
- Prompt quality has a ceiling; the models absorbed most of what prompt tricks used to buy.
- The durable work moved to context, tools, memory, and evaluation. The parts around the model.
- Agents fail on plumbing, not reasoning: bad retrieval, silent tool errors, no stopping condition.
- If you cannot say what "done" means for a run, you do not have an agent. You have a chatbot with ambition.
There is a specific kind of disappointment that hits about six weeks into building with an LLM. The demo was magic. The prototype was fine. Then real inputs arrived and the whole thing turned into a coin flip, and the reflex is to go back and rewrite the prompt. Again. With more capital letters.
That reflex is the last surviving habit from the prompt-engineering era, and it is now the single biggest waste of engineering time in applied AI.
The prompt got commoditized on purpose
Model vendors spent two years absorbing the tricks. Chain of thought, role-play framing, "think step by step," structured output coaxing. Most of it is now either default behaviour or a first-class API parameter. What used to be craft is a checkbox. That is a good outcome and it means the leverage moved.
By 2026, most serious teams describe their work as context engineering instead: deciding which tokens earn a place in the window, and in what order. Sourcegraph frames it as choosing what the model sees rather than how you ask (Sourcegraph (opens in a new tab)). That reframing sounds subtle and is not. It moves the problem from writing to retrieval, ranking, compaction, and state. Problems with actual engineering answers.
Where agents actually break
Read enough production postmortems and the failure modes rhyme. Almost none of them are "the model could not reason."
- Silent tool failures. A call returns an error string, the model treats it as data, and the run continues on a fiction for six more steps.
- Context rot. Fifteen turns in, the useful instruction is buried under transcript and the model starts optimizing for the wrong thing.
- No stopping condition. Nobody defined done, so the agent either quits early or loops until the budget dies.
- Memory that is really just a log. Everything is stored, nothing is retrieved, and the model behaves like it met you for the first time this morning.
- Zero observability. When it goes wrong there is no trace, so the fix is a guess wearing a lab coat.
What a system has that a prompt does not
The parts list is boring, which is why it took so long to get taken seriously.
- A typed contract. Inputs validated, outputs schema-checked, failures raised instead of narrated.
- Retrieval you can inspect. Not "we vector-search everything", a ranked, logged, budgeted set of documents with a reason for each.
- Durable memory with a decay policy. What is worth remembering next week, and what should have been forgotten yesterday.
- Tools with real error semantics. A tool that can fail loudly is worth three that fail politely.
- Evaluation. A golden set, a judge, and a number that moves. Without it, every change is vibes.
- Cost accounting per run. Tokens are a budget line, not an abstraction.
None of that is model work. All of it is software engineering, which is the actual news of the last eighteen months: the interesting part of AI became normal.
How this shows up in ClawLI
Our writing agents are unremarkable at the prompt layer. The system prompts are short. What surrounds them is not: a character profile that carries the operator’s voice and audience, a memory layer scoped per organization, tools that read real CRM and analytics state, a cost event written for every run, and a gate that hard-blocks generation when the brief lacks proprietary detail.
When output quality drops, we almost never touch the prompt first. We look at what the model was handed. Nine times out of ten the fix is upstream, a stale profile, a retrieval that returned three near-duplicates, a tool that silently returned an empty array.
The best prompt in the world, handed the wrong context, produces a confident wrong answer faster than a bad prompt does.
The practical migration
If you are still living in the prompt file, the move is not a rewrite. It is three weeks of instrumentation. Log every run with inputs, retrieved context, tool calls, and tokens. Build a golden set of thirty real cases with known-good outcomes. Then start changing things and watch the number.
You will find that half your prompt is dead weight, that your retrieval is worse than you assumed, and that the model was never your bottleneck. That is the point. Related reading: context engineering, and what it replaced and evals or it did not ship.
The full prompt library we run internally is documented on my site (opens in a new tab), and the client work that stress-tested it happens at the TechTide AI studio (opens in a new tab). Steal the structure, not the wording. Your business context is the part that matters.
“A prompt is a sentence. A system is a sentence, plus everything that decides whether the sentence was any good.”
Questions people actually ask
- Is prompt engineering dead?
- As a job title, largely. As a skill, it is now table stakes. Like knowing SQL. The differentiated work is in what surrounds the prompt: context selection, tools, memory, and evaluation.
- What is the difference between a prompt and an agent?
- A prompt produces one response. An agent runs a loop: it plans, calls tools, observes results, and decides whether it is finished. The loop is what needs engineering. Stopping conditions, error handling, budgets, and traces.
- Do I need a framework to build agents?
- No. Most production agents are a while loop, a tool dispatcher, and good logging. Frameworks help with orchestration and observability; they do not solve context or evaluation, which is where the real difficulty lives.
- How much context should I send?
- As little as answers the question. Longer windows made everyone lazy, and long-context recall is uneven in the middle of a window. Ranked, compacted context beats a full dump at both quality and cost.
Sources & further reading
Written by
Alex Cinovoj, Founder, TechTide AI
Alex Cinovoj builds ClawLI at TechTide AI, a fleet of specialist agents that research, draft, schedule, and grade LinkedIn content with a human approving every send. He writes about what breaks when you put agents in front of a real audience.
Keep reading
- AI agentsAutonomy Levels for Content AgentsA five-level framework for deciding how much your agents get to do on their own, and how they earn the next level.
- AI agentsMemory Is the MoatWhy your AI still writes like a stranger after six months, and what a memory layer has to do to fix it.
- Go-to-marketAfter the Crackdown: Operating a LinkedIn Program That Survives EnforcementPlatform enforcement caught up with automated presence in 2026. Here is how to run an AI-assisted LinkedIn program that does not depend on staying invisible.