Artificial intelligence applied to web development crossed, in eighteen months, the line separating demonstration from production tooling. This guide sets out what actually holds in production, what remains fragile, and the architectural decisions that separate a workable chain from an expensive prototype.
The landscape in 2026: three distinct layers
Confusing these three layers is the most common source of error in projects. They share neither constraints, nor costs, nor risks.
| Layer | What it covers | Maturity |
|---|---|---|
| Code-writing assistance | Completion, refactoring, test generation inside the editor | Established — measurable gain, low risk |
| Tool-equipped agents | A model reading from and writing to real systems through declared tools | Usable, under strict conditions |
| AI features in production | A model in the end-user journey | Demanding — cost, latency and moderation to control |
The protocol that unlocked agents
Before the Model Context Protocol was adopted, every connection between a model and an external tool was a proprietary script to maintain. MCP standardised that layer: a server exposes tools, any compatible client consumes them. The protocol became an open standard, handed over in late 2025 to a neutral foundation, and public registries now list several thousand servers.
The 28 July 2026 specification stabilised what was missing to build on top: a stateless core, OAuth 2.1 authentication for remote servers, and an announced deprecation policy.
What an MCP server changes in practice
- The agent fetches its own data instead of receiving it pasted into a prompt.
- Permissions are carried by the server’s authentication, not by trust in the model.
- One server serves several agents and several projects.
Running agents in production
This is the step separating experimentation from a workable system. Practitioners running automated chains converge on a small set of rules.
The automatable boundary
Anything mechanical and verifiable goes into automation; anything committing a signature, money, or a hard-to-reverse action stays on a human desk. This line is not a moral precaution: it is what bounds the worst-case outcome.
The five recurring failures
- Silent failure — a credential expires, the API returns 401, the chain “succeeds” while producing nothing. Every run must emit a readable report.
- Blocking at the source — an application firewall refuses the crawl, and the agent invents what it could not read.
- TLS inspection in cloud environments, whose obvious “fix” — disabling certificate verification — is a security incident.
- Prompt drift when several sessions write to the same reference document.
- Duplicates, in the absence of a stable deduplication key checked across all statuses.
Editorial agents in production: what holds and what breaks covers these points on a real case, with a trigger configuration and the sub-agent split.
The real cost, rarely stated
The dominant cost item is not token consumption but plumbing: sources, deduplication, logging, dry runs, error recovery. Return on investment is decided in the engineering around the model, not in the model. A project that budgets the API without budgeting supervision ends up with a chain nobody dares leave running.
An automated chain is a production system. It requires monitoring, logging, guardrails and human responsibility at the end — exactly like the rest of the infrastructure.
Choosing: assistance, agent or feature
| Need | Suitable answer | To avoid |
|---|---|---|
| Speed up code writing | Assistant inside the editor | Building an agent for it |
| Automate a repetitive, verifiable task | Scheduled agent with declared tools | Letting it publish or spend without validation |
| Enrich a user-facing product | Application integration, with caching and fallback | Calling the model synchronously with no cost ceiling |
| Explore an internal corpus | MCP server on the source of truth | Copying the data into the prompt |
The questions to settle before starting
- What happens if the model is wrong? If the answer is “a human sees it before it becomes visible”, the project is viable.
- What is the source of truth? An agent without a structured source produces plausibility, not information.
- Who reads the report? Automation with no human recipient ends in silent failure.
- What is the spending ceiling? To be set before the first run, never after.
Articles in this section
- MCP settles down: what the 28 July specification changes
- Editorial agents in production: what holds and what breaks
This page is updated as articles are published in the AI for the web section.