Context Engineering: The Skill That Makes or Breaks AI Agents

For the past two years, the conversation around AI productivity centered on prompt engineering — the art of crafting the perfect instruction. "You are an expert developer. Write a function that..." That approach works for single exchanges. But AI agents don't operate in single exchanges. They work through multi-step workflows, maintaining state across dozens of tool calls, file reads, and decisions.

That's where context engineering comes in. As Philipp Schmid from Google DeepMind puts it:

The distinction matters. Prompt engineering optimizes the question. Context engineering optimizes everything the model knows and can access when the question is asked — including retrieved documents, conversation history, available tools, and guardrails. According to a survey by DataHub, 82% of IT and data leaders say prompt engineering alone is no longer sufficient, and 95% consider context engineering important for building reliable AI agents.

Salesforce named context engineering one of the top eight AI agent trends shaping 2026, alongside deterministic guardrails and autonomous workflows.

Despite the crowded market, modern AI agent systems share a surprisingly similar architecture. Whether you're using Claude Code, Cursor, or Copilot, under the hood you'll find the same building blocks: slash commands (shortcuts for common tasks), skills (reusable capabilities), plugins (third-party integrations), hooks (automated triggers that fire on events), and MCP servers — the connective tissue that links agents to external data sources.

MCP — the Model Context Protocol, developed by Anthropic and now an open standard — acts as a universal connector between AI models and the outside world. Think of it as the USB-C port for AI agents: a standardized way for any model to plug into any data source or tool. IBM's architecture research shows MCP enabling everything from single-agent setups to complex multi-agent orchestrations.

Every piece of context an agent consumes has a price. Context is measured in tokens — the basic units of text that language models process. More context means more tokens, and tokens cost money.

Here's the paradox of 2026: per-token prices have plummeted, yet organizations' AI bills keep climbing. Why? Because agents are getting more ambitious. A coding agent that once edited a single function now restructures entire codebases — and each step consumes context. The Stevens Institute of Technology describes this as context overflow: agents accumulate so much history that they either forget their original instructions or blow past the token limit entirely.

There's an even more insidious problem than cost. Research from Redis and Chroma has documented a phenomenon called context rot — measurable performance degradation as the input context grows, even when the model's context window isn't close to full.

Some users report noticeable quality drops after using just 20% of the available context. For coding agents working on complex projects, this means the agent that started strong may produce increasingly sloppy output as the session drags on. As MindStudio notes, this is particularly acute in AI coding agents, where long sessions are the norm.

The solution isn't bigger context windows — it's smarter context management. Context engineering provides the frameworks: session resets, context compression, checkpoint-based workflows, and selective retrieval that keeps only relevant information in play.

The most interesting development in AI coding agents isn't the latest hyperscaler release — it's the growing ecosystem of open-source, model-agnostic tools that let developers choose any model, run locally, and control their data. (For a deep dive into 20+ CLI-based agent systems, see The Rise of CLI-Based AI Agents.)

OpenCode supports over 75 LLMs through a single interface, routing tasks to local models, open-weight models, or proprietary APIs depending on the job and the organization's data requirements.

Hermes (Nous Research) introduces a self-improving loop: the agent creates skills from experience and can delegate to specialized external agents like Claude Code or Codex.

Pi (Earendil Works) ships with just four core tools and a sub-1,000-token system prompt, extending itself through TypeScript extensions, markdown skill files, and prompt templates. Its modular architecture has become the foundation for other agents, most notably OpenClaw.

These tools share a common thread: they decouple the agent from any single model vendor — a practical application of context engineering at the infrastructure level. Proprietary agents like Claude Code and Cursor still lead on SWE-bench Verified, but the gap is narrowing fast, and for many teams, model flexibility matters more than a few benchmark percentage points.

For European organizations, the AI agent boom raises a pressing concern: data sovereignty. The US CLOUD Act compels US-headquartered companies to hand data to American authorities, regardless of where that data is physically stored. GDPR Article 48 explicitly states that a US court order alone is not a lawful basis for transferring personal data out of the EU.

As of May 2026, the EU is actively considering rules that would restrict member governments' use of US cloud providers for sensitive data. For AI agents that process code, documents, and business data, this creates a real compliance challenge — especially since agent tool calls and sub-agent delegations can cross borders automatically.

A growing class of open-weight models — models whose trained parameters are publicly available, so you can run them on hardware you control — offers a way out. Four names now define that field: GLM 5.3, Kimi K3, DeepSeek V4 and Qwen 3.8.

One caveat before the numbers. SWE-bench Verified, the coding benchmark quoted throughout this debate, has stopped separating the leaders: scores now cluster near the ceiling, and OpenAI has stopped reporting it altogether, citing contamination by training data. The tests that still discriminate are Terminal-Bench (real tasks in a live terminal) and DeepSWE (engineering work spanning hours).

DeepSeek V4 Pro holds the headline result: 96.4% on SWE-bench Verified, the best of any open-weight model and within a point of the overall leader. It ships under a permissive MIT licence — but at 1.6 trillion parameters, hosting it yourself is a serious infrastructure project.

Kimi K3 is the largest open model ever released at 2.8 trillion parameters, with a million-token context window. It leads the open field on both Terminal-Bench (88.3) and DeepSWE (67.3), and Moonshot is refreshingly blunt that it still trails the best proprietary models.

GLM 5.3 is the pragmatic choice. At 753 billion parameters it needs a fraction of the hardware, yet it beats Claude Opus 4.8 on terminal-based coding (88.2 vs 85.0). Note that its licence now requires a security review for very large providers.

Qwen 3.8 matters for reach: its 27-billion-parameter release runs on a single high-end GPU while still scoring competitively on SWE-bench Pro — frontier-adjacent capability without a data-centre budget.

• Context engineering has replaced prompt engineering as the critical discipline for building reliable AI agents. It's about what the model sees, remembers, and can access — not just what you ask it.
• Context rot is real: AI agent quality degrades as sessions grow longer, even well before the context window fills up. Smart context management is essential.
• Token economics is a strategic concern: Cheaper per-token pricing masks rising total costs as agents take on more complex, multi-step workflows.
• Open-weight models have closed the gap: GLM 5.3, Kimi K3, DeepSeek V4 and Qwen 3.8 now rival the best proprietary models on coding benchmarks — DeepSeek V4 Pro scores 96.4% on SWE-bench Verified, within a point of the overall leader — and all four can be self-hosted on European infrastructure, turning data sovereignty from a compromise into a realistic choice.
• MCP is becoming the standard connector: The Model Context Protocol is emerging as the universal plug for connecting AI agents to tools, data, and each other.

The organizations that master context engineering won't just build better AI agents — they'll build ones they can actually trust in production.