Author: Kartik (@code_kartik)
Source: Original article on X
In July 2025, Manus published one of the most quietly important numbers in agent engineering. Their average input-to-output token ratio in production: 100 to 1.
For every token an agent generates, it reads a hundred. The model isn't really generating anymore it's reading, again and again, growing its context with every step.
Whoever wins on what goes into that context and what gets paged out wins the agent product. Regardless of which model they're using.
This discipline now has a name. It's called context engineering, and it has quietly replaced prompt engineering as the most important skill in production AI.
What it actually is
Anthropic defined it in September 2025: "context engineering refers to the set of strategies for curating and maintaining the optimal set of tokens during LLM inference." Their framing of why it matters is worth quoting every token forces "n² pairwise relationships" in the model's attention. There is an attention budget, and every irrelevant token depletes it.
Walden Yan, co-founder of Cognition (Devin), said it more bluntly in June 2025: "Context engineering is effectively the #1 job of engineers building AI agents."
Andrej Karpathy publicly endorsed the term the same month. Six months later, Anthropic, OpenAI's Codex Frontier team, Manus, and LangChain had all converged on the same answer.
Bigger context windows make agents dumber
This is the contrarian part. Vendors compete on advertised context length 1M, 2M, 10M tokens like it's a feature. The data says otherwise.
In July 2025, Chroma published "Context Rot," testing 18 frontier models including GPT-4.1, Claude 4 Opus and Sonnet, Gemini 2.5, and Qwen3. The headline finding: "models do not use their context uniformly; their performance grows increasingly unreliable as input length grows."
Even on simple retrieval, performance degrades non-uniformly. A single distractor reduces accuracy. Logically-organized haystacks perform worse than shuffled ones. On a memory task with 113K-token full prompts versus 300-token focused prompts, every model performed dramatically better with less.
OpenAI's own MRCRv2 benchmark confirms it. Their previous-generation model dropped from ~88% accuracy at 256K tokens to ~37% at 512K-1M. The window grew. The reasoning collapsed.
The race to a 10M-token window is a marketing story. The real race is being run on cache hit rates and compaction strategy.
The four moves the winners use
Reduce. Anthropic's context editing automatic clearing of stale tool calls improved performance by 39% over baseline and reduced token consumption by 84% on a 100-turn web search evaluation. Claude Code auto-compacts at ~95% capacity. Cursor, Amp, and OpenCode all ship variants.
Offload. Manus treats the filesystem as "the ultimate context: unlimited in size, persistent by nature, and directly operable by the agent itself." OpenAI's Codex Frontier team learned the same lesson the hard way they tried a giant AGENTS.md and watched it rot. Their fix: "treat AGENTS.md as the table of contents, not the encyclopedia." A short ~100-line map; the actual knowledge lives in a structured docs/ directory the agent navigates on demand.
Isolate. Anthropic's research multi-agent system, with separate context windows per sub-agent, "outperformed single-agent Claude Opus 4 by 90.2% on our internal research eval." The architectural reason: each sub-agent gets a fresh, focused window instead of competing for attention in one mega-context.
Cache. Manus calls KV-cache hit rate "the single most important metric for a production-stage AI agent." The math: cached input tokens cost 10x less than uncached. An agent looping 50 tool calls with a 10K-token system prompt sends 500K tokens of identical prefix. Without caching, that's $1.50 per turn at Sonnet pricing. With caching, $0.15. KV-cache hit rate isn't a nice-to-have. It's the line between viable and unviable economics.
The Bitter Lesson kicker
Manus has rebuilt their agent framework four times soon to be five. Phil Schmid's December 2025 retrospective captured the lesson cleanly: "in the last six months, their biggest performance gains didn't come from adding complex RAG pipelines or fancy routing logic. The gains came from removing things."
They removed complex tool definitions in favor of general shell execution. They removed management agents. They killed their todo.md after finding it wasted ~30% of tokens. As models improved, the harness shrank.
Peak Ji's framing: "In the rising tide of model power, build speedboats, not pillars."
What this means for your team
ZipRecruiter now lists "context engineer" as a US job title with an average salary of $101,752. Gartner formally defined the role in 2026 and forecasts that by 2028, "context engineering features will be built into 80% of software tools used to build AI applications."
Lance Martin, who wrote the canonical taxonomy for context engineering at LangChain, was hired by Anthropic.
The signal is clear. If you're hiring "prompt engineers" or "AI engineers" without measuring KV-cache hit rate, compaction strategy, or context utilization, you're optimizing the wrong thing.
The model gives you intelligence. Context engineering decides whether that intelligence reaches your user.
Build accordingly.