The context window and compaction, multi-agent failure modes and error propagation, cross-session coherence, and confidence, sampling, and provenance.
Context degrades as it scales and as relevant information gets scattered across a long transcript — a model can technically "see" everything in its window and still perform worse than it would with the same information presented compactly and close together.
Token budget is a resource to allocate deliberately, not fill up: keeping headroom matters because a nearly-full context window is itself a failure mode, leaving no room for the model to actually reason, plan, or incorporate a large tool result without something being pushed out or truncated.
Compaction (/compact in Claude Code) summarizes older parts of a conversation to free up space. What survives compaction is what was written into a persistent file (like CLAUDE.md); an instruction that only ever existed as a message in the conversation is exactly the kind of thing compaction is free to summarize away — which is why durable rules belong in configuration, not in a one-off chat turn.
Full breakdown & examples → 5.1 The Context Window
When a subagent fails, returning structured error context lets the coordinator make an intelligent recovery decision instead of guessing. Good error context has four parts: the failure type (transient — timeout or rate limit, may succeed on retry; validation — bad input, needs a fixed query; business — a rule violation requiring escalation or an alternative path; or permission — access denied, unretryable without a change in authorization), what was attempted (the specific query or action, stated concretely rather than as "it failed"), partial results gathered before the failure, and alternatives the coordinator could still try.
Two anti-patterns recur constantly: silent suppression, where an empty or failed result is treated as if it were a valid empty answer, so the coordinator never realizes recovery is needed at all — the worst of the two, because it prevents any recovery from ever being attempted; and workflow termination, where one subagent's failure kills the entire pipeline, discarding perfectly good partial results other subagents already produced.
The right split of responsibility: subagents handle their own transient retries locally (a timeout doesn't need to escalate all the way to the coordinator), while the coordinator reserves its attention for failures that need a genuinely different strategy — a different data source, a narrower query, or a human escalation.
Full breakdown & examples → 5.2 Multi-Agent Failure Modes
Work that spans multiple sessions needs a place to persist knowledge outside any single context window — external memory, written summaries, or artifacts the next session can load back in, rather than relying on everything staying in one continuous transcript.
Coherence across a long-running or resumed task means the agent picks back up with the same understanding of state it had before — the same goal, the same constraints, the same progress — which only works if what got persisted actually captured the parts that matter, not just the most recent few messages.
Full breakdown & examples → 5.3 Cross-Session Knowledge and Coherence
A model's own expressed confidence is a signal, not a guarantee — it can be miscalibrated, especially on questions outside common patterns in training data, so system design shouldn't treat a confident-sounding answer as equivalent to a verified one.
Sampling parameters (like temperature) trade determinism for variety: lower settings produce more consistent, repeatable output, which is usually what you want for a structured-extraction or classification task, while higher settings suit creative or exploratory generation where variety is the goal.
Provenance — tracking where a claim in the output actually came from, a specific search result, a specific document, a specific tool call — is what makes an agentic system's output auditable after the fact, and what separates a system you can debug from one where a wrong answer is a mystery.
Full breakdown & examples → 5.4 Confidence, Sampling, and Provenance