An AI demonstration takes minutes. A business task lives for weeks or months, touches several systems, and ends not with an impressive answer but with a changed state.
I have applied an agentic approach to hundreds of business tasks. Publishing an exact number without an anonymized export would add little value. The useful evidence is the set of recurring reasons why work looks complete too early.
1. A vague task produces a persuasive adjacent result
"Improve the service" has no boundary. An agent may make a reasonable change and still fail to solve the user's problem.
A working brief names the observable result, constraints, and method of verification. For example: remove a configuration mismatch, preserve backward compatibility, and prove the behavior in a test environment.
The clearer the boundary, the less room remains for an attractive substitution.
2. More context is not always better
Without context, an agent repeats rejected decisions. Given a huge pile of documents, it may select an obsolete rule or lose the objective in detail.
The goal is sufficient context, not maximum context: current requirements, accepted decisions, interfaces, and facts for the active step. Everything else should remain available without being loaded by default.
3. The author should not be the sole verifier
A model tends to continue its own account of events. If it writes code and is then asked whether everything is complete, it often checks the internal coherence of the solution rather than the real state.
Verification must be independent and happen at the right boundary. For code that may be a test environment. For a migration, the database state. For publication, the public surface. For an article, fidelity to the author's intent.
The failure inside this series
The first edition of this series was technically orderly: 13 slides, 11 article pairs, a working deck, caveats, and tests. The articles, however, mechanically copied the presentation structure. A slide intended for a few minutes of spoken explanation became a separate article. To look complete, each thesis acquired transitions and repetition.
The package passed formal checks and still felt padded. This was the same class of failure as losing Arcanada's origin story in an earlier revision: the system satisfied a measurable requirement without preserving the meaning.
The correction was not stronger editing of eleven files. The model itself had to change. The presentation keeps 13 slides, while related ideas become seven arguments designed for reading.
What works in practice
Good agentic work leaves four things behind:
- a result with a clear boundary;
- evidence at that same boundary;
- the preserved reasoning behind the decision;
- an honest list of what remains unproven.
Hundreds of tasks do not prove that the method is error-free. They provide enough cases to identify recurring failures and turn them into operating rules.
AI-First becomes an engineering approach not when an agent writes a lot, but when speed remains subordinate to meaning, verification, and the human ability to stop and say: "We solved the wrong problem."