The full history is not a contract
Giving every agent the complete conversation, every document and every decision feels convenient, but makes it difficult to know which information actually influenced the output.
Role-specific work views
The orchestrator can build a context envelope for each task: required evidence, permissions, goals, constraints and references. Durable memory still exists; the agent receives a minimal view.
Less exposure, more auditability
Reducing sensitive context decreases privacy surface and makes execution easier to reproduce. It also prevents an auditor from seeing information it does not need to challenge an artifact.
More context can weaken the signal
Sending the entire history increases tokens, exposure and the chance that irrelevant information influences output. An agent reviewing a migration does not necessarily need commercial conversations or secrets from other projects. Context selection is part of task design.
Views by responsibility
Each role can receive a different envelope: objective, necessary evidence, authorized files, current decisions and references for retrieving more. This preserves global memory without turning it into an indiscriminate dump into every execution.
Retrieve on demand
Minimum context does not mean depriving the agent of information. It can query sources when a concrete question requires them. That retrieval leaves a trace of which material influenced the work and avoids loading data that may never be used.
Privacy by need
Reducing context also reduces exposure surface. An agent or provider sees only what its function requires. This complements permissions and tokenization by limiting volume before more specific controls are applied.
Audit becomes simpler
When the input package is bounded, reconstructing why an output could contain a claim becomes easier. Unlimited context makes provenance diffuse and turns debugging into conversational archaeology.
A useful test
One concrete way to put this idea under pressure is to give an agent only a minimum package, allow on-demand retrieval and compare utility, cost, exposure and traceability against a run with full history. The test should not ask only whether an answer appears, but which state remains, what evidence is preserved and whether another operator can understand why the system behaved that way. This turns an editorial principle into an observable property and exposes places where architecture still depends on invisible assumptions.
What this note does not claim
Minimum context does not mean hiding necessary information; it means justifying why each datum enters a task and enabling additional access when there is a concrete need. This distinction matters because a good practice stops being useful when it becomes a universal promise. The goal is to make one design boundary explicit so it can be discussed, tested and adapted to the domain while facts, inferences, permissions and decisions remain separate.