Skip to content

Context Discipline

Context discipline is the practice of managing what an AI agent reads, retains, and passes forward. It is the intake, retention, and handoff hygiene that prevents context bloat, reduces token costs, and keeps agents focused on their assigned task.

The Three Practices

Context discipline has three complementary practices:

Practice Question Asset
Reading discipline What should I read? context-discipline fragment
Compaction What should I keep? context-compaction fragment
Handoff What should I pass forward? context-handoff template

1. Reading discipline

Reading discipline governs what an agent reads before acting. It is the intake side of context management: read the minimum to answer the question, stop when you have enough, and never read "just in case."

The context-discipline fragment defines the rules: a context budget (max 10 files before a change), a priority order (task first, then files being modified, then types, then standards), and anti-patterns (no speculative reads, no re-reading unchanged files).

2. Compaction

Compaction governs what an agent retains after a phase boundary, a handoff, or a context-pressure event. It is the retention side of context management: preserve decisions, constraints, and evidence; drop execution trace, verbose output, and resolved questions.

The context-compaction fragment defines the compact-after triggers (phase boundary, handoff, drift, window full, retry, human gate) and the never-compact-away items (task, constraints, decisions, open questions, evidence, file paths, next action, pending handoffs).

The session-compaction template provides a structured format for mid-session compaction. The recovery-handoff template provides a structured format for compaction after a failure.

3. Handoff

Handoff governs what an agent passes to another agent or a future session. It is the transfer side of context management: package the essential state, decisions, and evidence into a structured document that the next agent can consume without re-deriving context.

The context-handoff template provides a structured format with evidence requirements, trace evidence entries, and a required-evidence checklist that the receiving agent must verify before proceeding.

Why context discipline matters

Without context discipline, agents exhibit predictable failure modes:

  • Context bloat — the agent reads more than it needs, filling the context window with irrelevant information and degrading performance.
  • Decision loss — the agent makes a decision, then forgets it, and re-derives it (or worse, makes a contradictory decision).
  • Constraint drift — the agent satisfies a constraint early, compacts it away, and violates it in a later phase.
  • Evidence erosion — the agent claims "tests pass" without preserving the test output, leaving the next agent unable to trust the claim.
  • Handoff gaps — the agent ends a session without recording what was done, what was decided, or what comes next.

Context discipline prevents all five failure modes by making the intake, retention, and transfer of context explicit and structured.

Relationship to other concepts

Concept Relationship
Token rent Token rent is the hard byte budget for the canonical library; context discipline is the behavioral practice that keeps agent sessions within their effective context window.
Minimality Minimality is the principle that every byte in the canonical library earns its place; context discipline is the same principle applied to session context.
Agent contracts Agent contracts define what an agent does; context discipline defines how an agent manages its working memory while doing it.
Harness principles The "Handoff and memory hygiene" principle is the harness-level expression of context discipline.
Trace/eval improvement loop Trace evidence feeds the improvement loop; context discipline ensures trace evidence survives compaction and handoff.

See also