A convincing answer is not enough
A language model can suggest an idea, draft an advertisement and divide a problem into steps. It does not automatically know which company facts are confirmed, who may spend the budget, whether a connection is still authorised or whether an external platform completed the requested action. Concealing those gaps behind confident prose creates an impressive demonstration, not an operational product.
A useful agent must distinguish knowledge, assumption, proposal and completed action. Each state needs its own evidence and its own boundary.
Context must belong to the project
The agent first needs authorised context: confirmed facts about the offer, audience, brand, connected channels and constraints. This cannot live safely in one endless conversation. It requires structure, versions, provenance and clear project ownership.
When sources conflict or become stale, the agent should expose the uncertainty by asking a question, proposing an update or limiting its conclusion. Silent guesswork turns convenience into risk.
Memory and evidence are not the same thing
It is useful for an agent to remember the chosen voice, working preferences and the order of a conversation. Conversational memory cannot, however, become the source of a business fact. A price, delivery condition, product limitation or legally significant claim needs a confirmed record rather than a sentence mentioned casually several weeks earlier.
This separation makes the system calmer to use. People can explore ideas without fearing that every thought will automatically become a brand rule. When a proposal genuinely belongs in the project knowledge base, the agent presents the change and asks for confirmation.
A proposal is separate from execution
When the user asks to launch a campaign, the agent should first prepare a proposal: goal, segment, messages, creative, budget, stopping conditions and expected service expenditure. Deterministic controls then verify permission, limits, approval policy and idempotency.
Only after approval does an execution job exist. Its outcome is checked against the external system’s response and written to an audit trail. The sequence is more deliberate than an instant button, but it protects against duplicate launches, the wrong project and invisible spending.
A working agent must be able to fail honestly
A provider may be unavailable, a model may violate the required format, an advertising account may revoke access, or the evidence may be insufficient. A real system distinguishes those causes and offers the next safe step. It does not replace failure with a prewritten success, nor does it quietly move to an expensive fallback while hiding the cost change.
Cost and quality need continuous observation
Even a good route changes over time: a provider updates a model, price rises, latency grows, or writing quality in a particular language shifts. Model choice therefore cannot be embedded permanently in marketing copy. It belongs to a release configuration tested against a representative evaluation set.
For every attempt, the system records the model, latency, expenditure, error and fallback use. The customer receives a clear final cost, while the product team gains evidence for a routing decision: where an economical model is sufficient and where a demanding task genuinely justifies a stronger one.
What good interaction feels like
In simple mode, the user sees a concise conversation, a clear proposal, the cost, the required approval and the verified result. In professional mode, the same operation opens to reveal sources, versions, attempts, parameters and the audit record.
The agent interface can therefore remain calm and simple while the system behind it remains rigorous. Simplicity is achieved by organising rules well, not by removing them.