Smarter AI Models Are Leaving Your Agentic Workflows Behind — What Anthropic's New Context Engineering Rules Mean for CX

We deleted the rules and nothing broke

On July 24, Anthropic published a technical deep-dive on "context engineering" — the art of assembling system prompts, skills files, and tool descriptions that shape how Claude behaves across many requests. The headline finding is stark: for Claude Opus 5 and Fable 5, the team removed more than 80 percent of the system prompt that powered Claude Code, with "no measurable loss on our coding evaluations."

Why? Because older models needed rules. They were bad at following complex instructions, so developers compensated with explicit guardrails — step-by-step procedures, ALL-CAPS warnings, reminders not to skip step 9B. The resulting prompts started to resemble furniture assembly instructions. Anthropic's new insight, highlighted by Opus Research analyst Amy Stapleton on July 31, is that smarter models work better when you let them use judgment instead of piling on constraints.

Anthropic puts it bluntly: Before, give Claude rules. Now, let Claude use judgment.

Why this matters for enterprise CX — right now

The agentic AI hype cycle has delivered a wave of customer service automation: AI agents that look up orders, check refund eligibility, decide whether to escalate, and actually execute multi-step workflows. Most of these are built with a mix of deterministic backend logic and AI-powered judgment. The deterministic pieces — an API call to query order status, a rule-based refund check — don't change when you swap the model. But everything upstream of them might.

As Stapleton notes: "The model deciding whether to call the API might change. A more capable model might infer a value the old one would have stopped to ask for. It might also run redundant checks left over from a prompt written to babysit a weaker predecessor."

This is not theoretical. If your CX vendor upgrades from an older Claude model to Opus 5, your carefully tuned prompts — the rules, examples, capital-letter warnings — could start working against the system. The model might second-guess contradictory instructions, spend tokens reasoning through conflicts, or simply ignore safeguards it now finds patronizing.

The six paradigm shifts Anthropic identified

The new context engineering playbook isn't just about stripping tokens. Anthropic outlines six practical changes that every team building on Claude should internalize:

1. Give Claude rules → Let Claude use judgment. The old approach piled on "never do X" instructions. Now, higher-level guidance like "write code that reads like the surrounding code" works better.
2. Give Claude examples → Design interfaces. Examples constrain the exploration space. Instead, design your tools with expressive parameters and clear naming — the model figures out the rest.
3. Put it all upfront → Use progressive disclosure. Move detailed instructions into skills and tool descriptions that Claude loads on demand. Verification and code-review logic in Claude Code was moved into separate skills the agent calls selectively.
4. Repeat yourself → Simple tool descriptions. Stop duplicating instructions in both the system prompt and tool descriptions. Put tool usage instructions in the tool description only.
5. Memory in CLAUDE.md → Auto-memory. Claude now automatically saves relevant memories across sessions rather than relying on users to manually log preferences into markdown files.
6. Simple specs → Rich references. Markdown plan files gave way to HTML artifacts, test suites as specs, and "rubrics" that let Claude verify its own work against your quality standards.

The CX maintenance problem no one talks about

Opus Research makes a pointed recommendation: CX platforms need their own version of claude doctor — Anthropic's new diagnostic command that automatically flags bloated prompts, redundant checks, and outdated instruction patterns when you upgrade models.

The scary scenario? A team discovers that large portions of their "permanent" customer-service operation were actually just behavioral advice written for an older model — advice the new model no longer wants. The business logic should be stable. Prompt scaffolding shouldn't have to be, but in most enterprise deployments today, those two layers are hopelessly entangled.

The healthy long-term setup, Stapleton argues, separates the two layers: a stable definition of what a customer-service skill is supposed to do (policies, permissions, evaluation criteria) sitting apart from model-specific instructions. When you upgrade the model, you rewrite the instructions, not the business rules.

What it costs: Claude API pricing, August 2026

Since every CX team will ask the money question: here's what you pay for Anthropic's Gen 5 models today. Prices are shown per million tokens (MTok), with an approximate EUR conversion at today's rate (1 USD ≈ 0.92 EUR):

* Sonnet 5 introductory pricing through August 31, 2026. Standard pricing: $3/$15 per MTok input/output thereafter.

For most CX use cases — identifying that a customer is asking about an order and pulling the shipping status — you probably don't need Opus 5 at all. Haiku 4.5 at €0.92 per million input tokens handles intent detection well, sometimes paired with conventional speech recognition that doesn't burn tokens. As Stapleton notes, the model upgrade only really matters once the AI is running a multi-step agentic workflow where judgment and tool selection come into play.

European availability: AI Act and GDPR implications

Claude is fully available across the EU — including the Czech Republic, Slovakia, Germany, France, and every other member state — for both Claude.ai and the commercial API. Anthropic Ireland Limited provides services in the EU, which matters for GDPR compliance: data processing falls under Irish jurisdiction, the company's EU base.

For European companies deploying AI agents in customer-facing roles, the AI Act's classification system becomes relevant. An AI agent handling returns and refunds for a telecom operator likely falls under the "limited risk" category — transparency obligations (users must know they're talking to AI), but not the heavy conformity assessment required for high-risk systems. However, if your agent makes legally consequential decisions — denying insurance claims, rejecting loan applications — that crosses the high-risk threshold, and the model-upgrade issue gets more complicated: you may need to re-certify the system after a model swap.

The practical takeaway: European CX teams evaluating Anthropic's Gen 5 models should document their prompt migration as part of their AI Act compliance trail. If you strip 80% of your system prompt like Anthropic did, you want a record of what you changed and why.

From our own production experience

We run AI services in production here at ai-jarvis.eu — article generation pipelines that orchestrate multiple model calls, tool invocations, and quality checks. The lesson from Anthropic's context engineering post resonates directly. When we moved parts of our pipeline from older models to Claude Sonnet and Opus, we saw instances where elaborate instruction sets — "do this first, then check X, never do Y" — actually caused the model to reason itself into circles. The fix wasn't better prompts; it was shorter prompts that trusted the model's native judgment.

One specific finding worth flagging: our pipeline runs with a system prompt that includes a long writing guide. Anthropic's observation that "conflicting messages in a single request" force the model to expend extra reasoning time before acting is something we've seen firsthand. A prompt that simultaneously says "do not add comments" (from the system prompt) and "document this function" (from the user) creates friction that didn't exist with simpler models simply because they weren't capable of noticing the contradiction.