The AI Brief #54 ai-agents-production llm-reliability ai-security smb-automation operational-risks

AI agents in production: it's infinitely more complex than anyone admits

Rodrigue Le Gall | | 3 min read

The conversation around AI agents has shifted. Gone are the days of marveling at an autonomous loop capable of completing a task. In 2026, teams that have actually deployed agents in production are discovering the reality: the complexity isn’t in building the model—it’s in operational stability.

The real problem surfaces when an agent needs to send messages, approve purchases, move money, or make decisions that commit your company. At that point, the question stops being “can the AI do this?” and becomes “can we prove who made this decision?”

Three concrete obstacles are blocking deployments:

  1. The absence of verifiable identity: when an agent acts on behalf of your company, how do you prove its integrity? There’s no satisfactory answer today.

  2. Reliability of structured output: getting an LLM to produce valid JSON matching a strict schema every single time is a nightmare. Teams in production report non-negligible error rates even with the best models.

  3. Memory fragmentation: an agent that loses context halfway through, or that mixes up user sessions, creates an immediate compliance risk.

These problems aren’t bugs fixable by an update. They’re architectural challenges. Companies pushing agents without solving these three points discover their limits too late.

What this means for your business

For your small business: stop chasing full autonomy. AI agents shouldn’t decide alone. They should verify, summarize, propose—and you stay the decision maker. If you deploy an agent on a task without clear traceability of who did what, you create legal and operational risk. Real value lies in tasks where the agent completes 80% of the work (sorting, classification, synthesis, preparation), but you keep final approval. This reduces your workload without sacrificing control. Start there before dreaming of autonomous agents.


In brief

Security: how malicious images bypass AI code review

An attack called Ghostcommit exploits the vision capabilities of AI code review tools by hiding malicious code in PNG files. Multimodal models validate the code without seeing the danger. The problem: you’re trusting a tool that reviews nothing. Direct impact if you’re automating your code reviews.

Read source

Shared memory in AI: a systemic security risk

ChatGPT, Claude, and Gemini all store a user’s contexts in the same session. Result: if your grandmother opens an AI, discusses her allergies, then opens another conversation, the AI can mix sensitive contexts. For a small business, this means risk of customer data leakage if you integrate these tools without strict isolation.

Read source

Structuring LLM output: 3 months of production broken down

A production team tested every trick to get valid JSON: simple prompts, fine-tuning, constrained decoding. Result: even 70B models regularly fail to respect a strict schema. This directly affects any integration system that depends on reliable structured formats.

Read source

Superhuman perfects AI email drafts: almost usable

Superhuman’s new auto-draft feature generates email responses that require little to no editing. It signals that LLMs have finally mastered a daily use case without major mishaps. If you’re looking for AI automation without reputational risk, this is it.

Read source

Apple vs OpenAI: lawsuit over trade secret theft

Apple accuses OpenAI of coordinated industrial espionage through its hardware CTO. Beyond the media noise, this raises a concrete question: if Big Tech companies are accusing each other of data theft, how do you protect your business data when using their tools?

Read source

Get The AI Brief in your inbox

3x per week, the essentials of AI decoded for business leaders.

Subscribe

Take action

Ready to automate your repetitive tasks?

Discover what AI can concretely change in your business. In 2 hours, we identify your automation opportunities.

Free AI Checklist

10 processes to automate in your business

Download PDF