I stopped prompt engineering. The work moved to the harness.
Essay, 2026-07-26
Models got smart enough to infer. The emerging craft is context and harness engineering: rich context, then memory, markdown, skills and MCPs around the agent.
The way we interact with AI is changing. For a long time the craft was prompt engineering: spell every step, hedge every edge case, write instructions as if the model would only do exactly what you typed. I have not worked that way for about a year. Somewhere around the GPT-5 lineup and Anthropic's Sonnet 4 generation, the models got smart enough that the old level of spelling-it-out stopped paying for itself. They can infer. The useful work moved elsewhere.
What matters now is context, then the harness you put around your agent: memory, markdown, skills and MCPs. That emerging field is context and harness engineering. You are not polishing a single magic sentence. You are building the room the model works in.
Less prompt, more context
My favourite game lately is the opposite of prompt engineering. How much rich context can I give an agent, and how little prompt can I get away with, and still land real work? Not as a stunt for a demo. As a habit. When the model can read the repo, the prior decisions, the tools, and the definition of done, a short ask is often enough. When it cannot, no amount of carefully worded instruction fully covers the gap.
Anthropic's own writing on the newer Claude models points the same way. They cut a large share of Claude Code's system prompt without losing ground on their coding evaluations, and shifted toward judgement, progressive disclosure, and tools that load when needed. The industry is catching up to something builders have been feeling in their hands: over-specified prompts start to fight the model instead of helping it.
The harness is the product
The harness is memory, markdown, skills and MCPs. Memory that keeps what mattered from the last session. Markdown that holds plans, voice and gotchas in a form the model can read cleanly. Skills and MCPs that decide what the agent can actually do: open a PR, run a check, talk to a browser, or only draft text in a box. That stack is what turns a clever chat into something you can trust with real work: hands-off agent runs, swarms you despatch and review, tasks that keep going in the cloud while you do something else.
Tools like Claude Despatch and Remote Control, and agents like OpenClaw and Hermes, are part of that shift. They pull the centre of gravity away from sitting down, opening a chat window, and grinding through a multi-turn conversation until the answer appears. The session is no longer the whole product. The environment is.
What this frees up
None of this means prompting disappears. You still have to say what you want. It means the scarce skill is no longer coaxing a model through every comma. It is deciding what context is load-bearing, then what belongs in the harness around the agent, and how you know when the work is good enough to ship.
That is the evolution I am living in day to day. Shorter asks. Richer context. A harness of memory, markdown, skills and MCPs. Agents that keep moving when the chat window is closed. The models got smart enough to meet us there. The interesting work now is building the room they work in.
More field notes