Compress for agents, not for humans

Most “compression” advice is written for humans: shorter sentences, fewer adjectives, a tighter narrative arc. Agents fail differently. They don’t get bored; they get diluted. Once the window fills with explanations a model already knows, soft opinions, and duplicate restatements, the load-bearing facts stop being load-bearing. The fix is to compress for agents: optimize value delivered per token, not how the note reads aloud.

That premise is the spine of Context Compression, and it shows up wherever I hand work between sessions: BRIEF.md for discovery and handoff shape, TLDR Research Ops for research that has to survive a fresh context. The methodology itself started earlier: the July 2025 Context Engineering notes and dual channels and rule layers are where “value per token” stopped being a slogan and became architecture. Agents deserve a real context format is the standards-facing sibling; this post is the methodology behind why those formats exist.

Attention budget

A context window is an attention budget, not storage. Tokens compete. If you treat the window like a warehouse, you’ll ship a shorter document that is somehow worse, because you cut the wrong things to hit a size target.

Value-per-token reframes the goal. Keep what changes decisions. Cut what only soothes a human reader. The metric isn’t compression ratio; it’s whether an agent with this note can act without a clarification round you already paid for once.

What must never be trimmed

Some content is sacred relative to any budget: decisions already made, interfaces the agent must honor, and constraints that make a “clever” solution invalid. Trim explanations, motivation, and repeated background first. Never trim those three just to look smaller.

An early decisions/interfaces/constraints-only model worked for specs and broke on research notes, plans, and chat. The generalization that stuck was a four-part shape any content type can map onto: identity, facts, relationships, and next actions. Different genres fill the slots differently; the slots stay stable so a downstream agent knows where to look.

Offline distill → attach

Don’t make the live session rediscover the project from raw docs. Distill offline (profile the source, extract structure, run ordered compression passes, validate), then attach the distilled note when the agent starts.

Offline is where the expensive passes belong. Live is where judgment belongs. If the agent is still reading the novel version of your architecture notes mid-task, you’ve spent attention on archaeology instead of the decision in front of it.

Structured handoff shape

Distillation alone isn’t enough if the handoff is shapeless. The next agent needs a predictable place for “what this is,” “what’s true,” “how pieces relate,” and “what to do next.” That’s why BRIEF.md and the agent-ready note formats keep the cut deliberate: discovery, scope, and handoff shape travel; storage and retrieval stay out of the standard.

Structured handoffs also make dual-channel work cheaper. A builder can return a pointer and a short summary; a reviewer can open the distilled artifact cold. The shared window stays small because the payload lives in a file with a known shape.

How this later became tooling

Once the methodology stopped being a personal checklist, it wanted tooling: ordered passes you can log, a validation path that fails closed instead of shipping a lossy note, CLI and service surfaces that apply the same rules without copying them into every prompt. Context Compression is that tooling layer. BRIEF.md is the discovery-and-handoff piece. Auto-research is what happens when the same discipline meets a research corpus that has to stay usable across sessions.

The claim I’m willing to make is modest. Teams that adopted agent-ready notes anecdotally reported fewer clarification rounds downstream (an impression, not a benchmark). The harder claim is architectural: if your agents keep asking for the same facts, you don’t have a prompting problem. You have an attention-budget problem, and the fix is distillation plus structure, not a longer prompt.