CCA-F Part 6: Context Management and Reliability (15%)

Part 6 of 7 of Claude Certified Architect — Foundations: a complete self-study tutorial. See all parts.

Domain 5: Context Management & Reliability (15%)

The smallest domain by weight, but a frequent source of subtle wrong answers.

Context windows

The current Claude family (June 2026):

Claude Fable 5 and Claude Mythos 5 also have 1M-token windows and were released 9 June 2026.

When a session approaches the context limit, Claude Code automatically compacts prior messages into a summary. Anything past the compaction is gone unless preserved by a tool call result you still have, a file you saved, or the summary itself.

Preserving critical information

Three patterns the exam favours:

1. Scratchpad file. Have the agent write important findings to a file (SCRATCH.md, NOTES.md). Reload as needed; survives compaction.
2. Subagents for token-heavy reads. The subagent reads 200 files and returns a 500-word summary. The parent's context never holds the 200 files.
3. Explicit "memory" tool calls. Save and retrieve facts via a tool that writes to durable storage (a JSON file, a vector DB, an MCP server). On critical workflows, do this rather than trust the context window.

Reliability patterns

The exam expects you to choose the production answer, which usually means:

• Cap retries (typically 2 or 3) and escalate. Never an unbounded retry loop.
• Idempotent operations. Especially for tools that write or send. The exam loves "the agent retried after a network timeout but the email was sent twice — what is the fix?" Answer: idempotency key on the tool call.
• Multi-pass review. For anything user-facing, draft → validate → optionally rewrite → emit.
• Graceful partial failure. When a subagent fails, the coordinator should proceed with what it has and flag the confidence drop — not abandon the entire workflow.

The escalation rule

When confidence is low or the request falls into a sensitive category, hand off rather than guess. The wrong exam answers are usually variants of "make Claude more confident" — increase temperature, add more examples, switch to Opus. The right answer is usually "this work shouldn't be done by the model; escalate."

<a id="scenarios"></a>

← Part 5 of 7: Tool Design and MCP Integration (18%) · All parts · Part 7 of 7: Exam Scenarios, Tips from People Who Passed, and All Resources →