03 — CONTEXT

Agents get what they need. No more, no less.

For every task, Altabric assembles a tailored context at runtime — the right documents at the right level of detail. The agent gets enough to do its job well, and no more. You can see exactly what was sent.

WHAT IT DOES

Context: how too much of a good thing can be bad.

The intuitive thing is to give the agent everything you have. The full corpus, every related document, the complete history of decisions, all of it. The agent must surely do better with more.

That's not what happens. Beyond a certain point, more context degrades the output — the agent loses focus, attends to the wrong things, repeats itself, contradicts material it saw two pages ago. Anthropic's own engineering team calls this context rot: as the number of tokens in the context window increases, the model's ability to accurately recall information from that context decreases. Context, they argue, must be treated as a finite resource with diminishing marginal returns. Their full piece is worth reading.

Altabric assembles context the other way around. For each task, the engine selects which documents to include, at what level of detail. The selection depends on the kind of task being run. The same document might be present in full for one task, in summary for another, and absent altogether for a third.

THE COMPRESSION LADDER

Five levels of detail. Each task gets the level that fits.

Documents flow into agent context at one of five compression levels. The level is chosen per document, per task — based on how central that document is to the work at hand.

FULL The complete document. Used when the task is grounded in the detail. e.g. designing against a spec RICH SUMMARY · ~500 WORDS Enough to ground the task without spending budget on every detail. the everyday default SHORT SUMMARY · ~150 WORDS Just enough to know what the document is and why it matters. background awareness HEADLINE Title and one-line description. peripheral mention OMITTED Not loaded at all. not relevant to this task

The ladder is applied per document, per task. A document that's central to one task may be omitted entirely from another.

TRANSPARENCY

What's in the context. What's not. Visible per task.

Every agent invocation comes with a context preview. You see which documents were loaded, at what compression, what was deliberately omitted, and why. No black box. No guessing what the agent was looking at when it made a call.

Context preview showing global, ancestry, parent, target, descendants, lateral, and entries sections — each with its compression level, document count, and token cost.

A real context preview from the engine. Each section names what was loaded, at what compression, and what was excluded. The token count is reported for visibility; the budget indicator alongside it is configurable and serves as a guide, not a hard limit.

EXAMPLE

What an agent actually sees.

REVIEWING A STORY FOR THE CUTOVER PLAN

One task. A precise selection of context.

An agent is reviewing a Story for the cutover plan workstream. The task is refinement and review, so the engine assembles the context accordingly.

The agent receives the full text of the parent Epic's Solution Design (the design it must align to) and the full text of the Story itself (the thing being reviewed). It receives the rich summary of the Theme PID one level up — enough to know the strategic frame without spending budget on it. It receives the short summary of two adjacent Stories that link in.

It does not receive the data migration architecture document, the change communications pack, or the test scripts. Those exist in the library, but they are not relevant to reviewing a cutover Story. The agent reads them when it is doing work where they matter — and only then.

TALK TO US

Tailored context. Controlled costs. Better output.

A demo runs against your delivery shape, not a generic one. We read every message. You will hear back within two working days.

Request a demo