Almost everyone who has seriously written code with AI has ridden the same emotional roller coaster.
We drew that roller coaster as a curve (see the cover image): the x-axis is how deep your practice goes; the y-axis is how you subjectively feel about it. Two rises, two falls, and finally a steady upward line. This is not one person's story — it is nearly everyone's. It is also the story we kept seeing, in ourselves and in our users, all along the road of building AVL Code.
Stage 1 · The Wonder of First Contact
Novelty, delight, anticipation.
The first time you toss in “write me a little tool to parse CSV,” runnable code lands in front of you within seconds. It even handles edge cases you had not thought of yet. That moment feels like someone pushing open a door for you, and behind it is a whole new world of “so this is possible.”
This stage is pure dopamine. You stay up all night feeding it every repetitive chore on your plate, telling everyone you meet that “this thing actually works.” The curve rockets to its first peak.
Stage 2 · Cursing at the Stupidity
The honeymoon is short.
Use it in earnest and disillusion arrives — disappointment, anger, doubt. It fabricates, with a straight face, an API that does not exist; it edits your perfectly fine function from bad to worse; the context it remembered one sentence ago is wiped clean by the next; you explicitly said “don't touch that file,” and that is exactly the file it touched. “How can it be this stupid? Infuriating!”
Here the curve falls to its first trough. Many people get off the ride at this point, leaving with a parting “it's nothing special.” But the fall itself is not an AI problem — it is the unavoidable road of breaking each other in: you have not yet learned to say things clearly, and it has not yet been given enough context. What those who get off miss is precisely the scenery further on.
Stage 3 · The Illusion of Omnipotence
If you make it through stage two — learning to break down tasks, feed clean context, and set enough constraints — the experience flips hard.
Confidence, inflation, illusion. “It can do anything, so I can achieve anything.” You start flinging it ever-bigger tasks as whole bundles, hitting Enter without so much as glancing at the generated diff; it says it needs to run a command, and you approve without lifting your eyes. Efficiency is frighteningly high; you feel like you have grown ten hands.
This is the curve's second peak, and its most dangerous point — because the steering wheel is slipping out of your hands without your noticing. We have seen the cost buried in this stretch too many times: an rm -rf nobody looked at closely, a suspicious sample “helpfully” executed as a script, a snippet of code that casually printed secrets into the logs. The high point of inflation is, more often than not, where the accident begins.
Stage 4 · The Fatigue You Can't Quit
When the illusion collapses, what follows is not sobriety but exhaustion.
Fatigue, dependence, burnout. You cannot leave it anymore — your hands have gone rusty, and without AI you hardly know how to begin; yet it wears you down over and over: tokens burn anxiously away, approval checkpoints stack up one after another, and once trust has broken even once, you look back to double-check its every step — until it is honestly more tiring than writing the code yourself. “Can't live without it, and it drains me dry.”
The curve bottoms out again. And this second trough is harder to cross than the first — because retreating to “no AI” is no longer possible. The only way out is through.
Stage 5 · Walking Together, Clear-Eyed
True maturity is the climb out of that stage-four trough.
Reason, collaboration, long-termism. You no longer expect it to do everything, and you no longer explode when it does something dumb; you treat it as a capable colleague who needs boundaries. You start setting rules: big tasks get a plan first, and nothing moves until you confirm; dangerous operations must pass through you; sensitive areas are off-limits, full stop; you embed it into your real workflow instead of letting it think in your place. “A tool — and a partner.”
This closing stretch of the curve has no more heart-stopping peaks — just a gentle, sustained, dependable upward line. And that is the entire reason we built AVL Code.
From day one we refused to make AI “a machine that charges off head-down at the press of a button,” and set out instead to build a colleague who knows self-discipline. In the product, that became five work modes — auto, plan, prepare, execute, assess: facing any sizable task, it first delivers a structured plan and todos, moves only after your confirmation, and assesses its own work when done. This “think first, act second” rhythm is precisely the antidote to stage three's “Enter without reading the diff.”
More important still is the steering wheel. Every tool call is approvable and auditable; the rm -rf check, the dangerous-injection blacklist, and the network allowlist never loosen in any mode. We also drew one hard line that cannot be lifted: samples/ is a read-only analysis zone where execution tools are force-disabled — even with “unrestricted mode” on, that line holds. Because people who do security for a living know best what stage-three inflation can cost. We give the AI freedom, but we do not hand over the wheel.
As for stage-four fatigue, what we care about most is sparing you worry and sparing you effort. Tokens burning anxiously? The AVL Delta high-efficiency transfer protocol sends only what each turn adds and never re-sends stable history; with prefix caching on top, it saves more tokens and gains speed. Sign in and a shared quota is assigned automatically — zero configuration, no wrestling with API keys. Dragged down by endless confirmations? Sessions are stored 100% locally and survive restarts; long-horizon tasks are handed to background subagents to run in parallel, so the main conversation keeps only clean progress lines instead of a screen drowned in intermediate logs. And the foundation is sovereign: SM4-GCM national-crypto encryption at rest, local-first, sample bodies never leave your machine, and the enterprise edition supports on-premises deployment.
None of this is a feature list dreamed up in the abstract. It was ground out of our own dogfooding, in one real long-haul task after another — from writing code to triaging a malicious sample, the same agent goes the whole way with you, and follows the rules the whole way too.
Maturity Is Not Deifying AI
Back to the curve. It does not end by soaring into the clouds; it ends as a steady upward line.
Maturity is not deifying AI — it is learning to collaborate with it, steadily. The wonder of stage one fades; the inflation of stage three bites back; only the clear-eyed companionship of stage five goes the distance.
If you have reached stage five yourself — or you are stuck at the bottom of stage two or stage four — come try an agent built, from its very first design decision, for “steady collaboration.” During the beta, the Zero starter plan is free for a limited time: download the desktop app, sign in with your phone number, and complete your first real session with zero configuration.
We are waiting for you at avlcode.cn — riding a donkey, with a steering wheel in hand.
AVL Code — the AVL security engine, with intelligence at your side. From the Antiy Landi team.