Context Engineering in 2026: What Changed After the Definitions

The definition, and the part most definitions leave out

Anthropic's engineering team defines context engineering as the set of strategies for curating and maintaining the optimal set of tokens during inference, including everything that lands in the window outside the prompt. Philipp Schmid's earlier framing, which did much to popularize the term in mid 2025, calls it the discipline of designing dynamic systems that provide the right information and tools, in the right format, at the right time. Both are accurate. Neither is the hard part. Sources: Anthropic Engineering; Philipp Schmid.

The hard part is subtraction. Almost every ranking explainer describes context engineering as assembly: gather the documents, add the memory, attach the tools, ship it. In a single-turn chat that is the whole job. In an agent running forty steps, removal is the discipline. Every tool result, every reasoning trace, every file the agent opened and no longer needs still sits in the window competing for attention with the instruction you care about. If your strategy has no answer for what gets deleted and when, you do not have one. This is why the term displaced prompt engineering rather than sitting beside it. A prompt is authored once. A context is running state an agent mutates every turn, and state that only grows is a leak.

The evidence that long context degrades, with the numbers

Context engineering rests on an empirical claim: performance is not uniform across input length. That claim is now measured. Chroma's July 2025 study evaluated 18 frontier models, including GPT-4.1, Claude 4, Gemini 2.5 and Qwen3, and found accuracy degrading non-uniformly as inputs grow, even on tasks as trivial as copying repeated words. One distractor reduced performance against baseline and four compounded it. The most counterintuitive result is that models scored better on shuffled haystacks than on logically structured ones. Sources: Chroma Research, Kelly Hong, Anton Troynikov and Jeff Huber.

The 2026 evidence is sharper because it comes from a safety-critical task with a clean control. In an arXiv paper published May 12, 2026, Sam Martin of the Anthropic Fellows Program and Fabien Roger of Anthropic measured how well models spot dangerous actions inside long agent transcripts. An obvious dangerous action inserted into a 100K token transcript was caught by Claude Opus 4.6 99.7 percent of the time. The identical action in an 800K token transcript was caught 69 percent of the time, and detection was worst when the target sat in the middle. That is not a retrieval failure. The information was present, unambiguous, and the model had been asked to look for exactly that thing. Length alone took thirty points off. Sources: Sam Martin (Anthropic Fellows Program) and Fabien Roger (Anthropic).

Google concedes a version of the same thing. Gemini reaches up to 99 percent accuracy retrieving a single fact from a very large context, and its documentation then states plainly that where you have multiple needles, the model does not perform with the same accuracy. It also names the tradeoff most vendor material avoids: retrieving the right information and controlling cost pull against each other. Sources: Google AI for Developers.

The four ways context actually fails

Drew Breunig's June 2025 taxonomy is still the most useful vocabulary for diagnosing a misbehaving agent. Context poisoning is when a hallucination or error enters the context and gets referenced repeatedly, so one bad inference becomes a premise. Context distraction is when the context grows so long the model over-focuses on it and neglects what it learned in training. Context confusion is when superfluous content in the window gets used and drags the answer down. Context clash is when new information or tools contradict what is already there. Sources: Drew Breunig.

The evidence he assembles for confusion is worth keeping. On the Berkeley Function-Calling Leaderboard every model performs worse when given more tools, and models sometimes call tools unrelated to the task. On the GeoEngine benchmark a quantized Llama 3.1 8b failed with 46 tools available and succeeded with 19. Same model, same task, and the only variable is how much of the window was spent describing capabilities it did not need. This taxonomy matters more than it did a year ago because agents now run long enough to hit all four failures in one session, and a forty-step agent has forty chances to poison its own context with no mechanism to notice. Sources: Drew Breunig.

The four operations everything else is built from

LangChain's July 2025 framing became the common vocabulary because it is exhaustive rather than clever. There are only four things you can do with context. Write it, saving information outside the window in scratchpads or persistent memory. Select it, pulling back only what this step needs. Compress it, summarizing or trimming what is already there. Isolate it, splitting work across agents so no single window carries everything. Sources: LangChain.

Anthropic's long-horizon guidance maps onto the same four with more detail: compaction, which summarizes history and reinitializes the window with the condensed version; structured note-taking, where the agent keeps external memory files that outlive the window; and sub-agent architectures, where specialists return summaries of roughly 1,000 to 2,000 tokens rather than their full transcript. That last detail is the one to copy. The value of a sub-agent is not parallelism. It is that the coordinator never sees the mess. Sources: Anthropic Engineering.

What actually changed in 2026: compaction became a parameter

Through 2025 every serious agent team wrote the same code: watch the token count, summarize the old turns, rebuild the window, hope the summary kept what mattered. In 2026 both major APIs document that behavior as a server-side feature, which is the difference between a technique and infrastructure. The two implementations differ in a way that tells you what each vendor thinks the problem is. Anthropic's context editing is surgical and rule-based: you declare what class of content to clear and when. OpenAI's compaction is wholesale and threshold-based: past a token count the server compacts the run and returns an opaque item carrying prior state forward. Sources: Anthropic, Claude platform documentation; OpenAI.

Your tools are context, and they are probably your biggest line item

The most common blind spot is that tool definitions are input tokens. Anthropic publishes the number: enabling tool use on Claude Opus 5 adds 286 tokens of system prompt before a single schema is counted, and on Claude Opus 4.7 it is 675. Schemas stack on top, and every tool result flows through the window on the way to the next step. Sources: Anthropic.

Anthropic's November 2025 write-up on code execution with MCP puts a figure on the compound version. In their worked example, an agent that loaded all tool definitions upfront and passed every intermediate result through the model used about 150,000 tokens. Restructured so the agent writes code that calls tools directly and loads only the definitions it needs, the same task used about 2,000 tokens, a 98.7 percent reduction. They give a case of the intermediate-result problem too: a two-hour meeting transcript passing through context twice can cost an extra 50,000 tokens. Sources: Anthropic Engineering.

Combine that with the Berkeley and GeoEngine results and the conclusion is uncomfortable for current integration culture. Connecting more servers to an agent makes it worse at using any of them. Exposing every tool because it might be useful is the most expensive context decision most teams make, and they make it once, by default, and never revisit it. Sources: Drew Breunig.

The unresolved argument at the centre of the field

Isolation is the one operation the field does not agree on, and most people quote only one side. On June 12, 2025, Cognition's Walden Yan argued against multi-agent architectures outright, because subagents misread their subtasks and make conflicting decisions with no visibility into each other's work. His principles: share context including full agent traces rather than individual messages, and recognize that actions carry implicit decisions, so conflicting decisions produce bad results. Sources: Cognition, Walden Yan.

One day later, Anthropic published the opposite result. Their multi-agent research system, with Claude Opus 4 leading Claude Sonnet 4 subagents, outperformed single-agent Claude Opus 4 by 90.2 percent on their internal research eval. They also published the cost: multi-agent systems use roughly 15 times more tokens than chat, and token usage alone explained 80 percent of performance variance on BrowseComp. Sources: Anthropic Engineering.

Our reading is that both are right, and the variable neither headline names is whether the subtasks write. Anthropic's case is research: read-only searches, no shared artifact, and conflicting decisions are impossible because subagents produce findings rather than changes. Cognition's case is code: subagents editing a shared codebase, where two locally reasonable decisions merge into one broken result.

The economics inverted, and most teams have not noticed

For most of 2025 the argument against large contexts was partly financial, because long-context requests carried a premium above a token threshold. That premium is gone on at least one major platform. Anthropic's pricing documentation states that Claude 4.6 and later include the full 1M token context window at standard pricing, and spells it out: a 900k-token request is billed at the same per-token rate as a 9k-token request. That removes the easy reason to keep contexts small and leaves only the real one. Do not fill the window because the model gets worse, not because the bill does. Anyone managing context purely as a cost line now has no forcing function, and the quality cost never shows up on an invoice. Sources: Anthropic.

That connects to something the Roundup tracked in its own public signal feed through July 2026, offered as our reading rather than anyone's published finding. A survey of 101 enterprises on agent orchestration found the platform choice broadly settled, most deployed agents still chatbot wrappers, and real-time fiscal control over token burn still the exception. The same week, the feed carried the argument for context-rich coding harnesses over grep-based retrieval. Those threads are one story. Teams are told to feed agents more context at the moment most cannot see what those agents already consume. Context engineering without token-level observability is guesswork with extra steps.

Stanford HAI's 2026 AI Index reports OSWorld agent accuracy climbing from roughly 12 percent to 66.3 percent while still failing about one attempt in three. Agents good enough to run long tasks are exactly the agents that accumulate enough context to degrade during them. Sources: Stanford HAI.

What to do, in the order that pays

Start by measuring, not fixing. Log input tokens per step for one real agent run and split the total into system prompt, tool definitions, retrieved content, history and tool results. Most teams are surprised by which of the five is largest, and the surprise is the point.

Then cut tool definitions before anything else. It is the cheapest change, it has the best evidence behind it, and it improves accuracy and cost at once. Audit which tools an agent actually calls, remove the rest, and load specialist tools on demand rather than upfront. Sources: Anthropic Engineering; Drew Breunig.

Next, adopt the platform compaction feature rather than maintaining your own, and set the trigger deliberately instead of accepting the default. Published degradation curves start well below 100,000 tokens on hard tasks, so if your agent reasons over its history, compact earlier and measure whether quality holds. Then give it external memory so compaction is survivable: if the only copy of a decision lives in the conversation, compaction can lose it, and if it lives in a file the agent can reread, compaction is free. Sources: Anthropic, Claude platform documentation; OpenAI; Chroma Research, Kelly Hong, Anton Troynikov and Jeff Huber; Anthropic Engineering.

Our AI observability guide covers instrumenting agents so these numbers exist at all, and our open-source AI agents guide covers the frameworks where most of this is configured.

Common questions

What is the difference between prompt engineering and context engineering? Prompt engineering optimizes the wording of an instruction, authored once and fixed. Context engineering optimizes the whole set of tokens present when that instruction is read, which changes on every step of an agent's run. Prompt engineering is a subset. It did not stop mattering, it stopped being sufficient once systems managed their own state. Sources: Anthropic Engineering; Philipp Schmid.

What is a context engineer, and is it a real job? The activity is real and the title mostly is not. This work sits inside AI, platform and applied research engineering roles as a named skill rather than a separate hiring line. Salary figures circulating for the title are too thin to quote at a source naming its methodology.

How do you learn context engineering? Read the primary write-ups from teams shipping agents, then measure your own. The Anthropic engineering post and the LangChain four-strategy framing are the most useful starting documents, and the Chroma study is the evidence underneath both. Paid courses exist, and none teaches more than instrumenting one real agent and seeing where its tokens go. Sources: Anthropic Engineering; LangChain; Chroma Research, Kelly Hong, Anton Troynikov and Jeff Huber.

Does a bigger context window remove the need for this? No. Detection of an obvious event fell from 99.7 percent at 100K tokens to 69 percent at 800K in Anthropic's own monitoring research, and Google's documentation concedes multi-needle retrieval does not hold single-needle accuracy. Bigger windows raise the ceiling. They do not flatten the curve. Sources: Sam Martin (Anthropic Fellows Program) and Fabien Roger (Anthropic); Google AI for Developers.

Limits and uncertainty

This guide describes context engineering as of August 13, 2026, in a field where the platform surface changes monthly, so check the linked API documentation before implementing against any parameter named here. Chroma's study and Anthropic's monitoring paper both come from organizations with a commercial interest in their conclusions, and both report on tasks of their own choosing. The 90.2 percent multi-agent result is an internal eval of undisclosed composition, not a public benchmark, and not independently reproducible. The Berkeley and GeoEngine figures are cited through Breunig's summary rather than read from the original leaderboards. Our reading of why Anthropic and Cognition disagree is editorial inference from two posts, not a claim either team made. The token-burn synthesis reads our own public signal feed, not original research. Where a number could not be verified at a source naming its methodology, including salary data for the context engineer title, we left it out.