Agent Context

Appearance

Agent ContextPortable context semantics for agent systems

Source refs, selection, budgets, assembly, injection, compaction, and missing-context facts for model, tool, runtime, UI, and peer-agent boundaries.

Read the specification
LLM full context

Explain what was available

Context surfaces and source refs make user messages, memory, knowledge, artifacts, tool results, browser state, and runtime state discoverable without merging them into one store.

Explain what was selected

Selection, ranking, omission, token budgets, and window records show why specific context entered a turn or task.

Explain what was injected

Assembly, injection, compaction, redaction, missing-context, evidence, and telemetry refs make final context reproducible and auditable.

Agent Context ​

Agent Context is a portable standard for the context layer in agent systems. It defines how available context is surfaced, selected, budgeted, assembled, injected, compacted, summarized, omitted, requested, redacted, and linked to evidence without replacing runtime schedulers, memory stores, retrieval systems, UI renderers, tool protocols, policy engines, artifact stores, or knowledge packages.

Use Agent Context when a product needs to answer: what did the agent know, where did it come from, why was this slice selected, what was left out, what was compacted, what was injected into a model or tool, and what context is still missing?

Draft standard for portable agent context surfaces, selection, budgets, injection, compaction, and missing-context facts.

Copyright © 2026