Part 10: Agent Loop Engineering

Part 9 established why an Agent needs a harness. Part 10 examines the runtime mechanisms that keep that harness useful under long contexts, large tool surfaces, restarts, mid-run user input, hung providers, repeated actions, and parallel calls.

Public-source snapshot: implementation details are examples from public Kocoro origin/main at commit 4ec6772, reviewed 2026-07-27. Treat the invariant as the lesson and re-check constants against current source.

Chapter map

Chapter	Topic	Question answered
34	From DAG to Agent Loop	Explains when a deterministic DAG should give way to an adaptive Agent loop, and what state a long-lived daemon must rebuild for every turn.
35	Context Compaction	Designs a staged context-compaction ladder that acts before overflow, verifies summaries, and retains an emergency recovery path.
36	Tool Result Budget and Spill	Shows how context offloading stores large tool results outside the prompt while preserving a useful preview and durable pointer.
37	Tiered Compression	Explains semantic compaction that preserves recent tool evidence and progressively reduces older results by age and value.
38	Deferred Tool Loading and Tool Search	Reduces MCP tool-schema bloat with token-budget and category-based deferral, a warm set, and model-visible tool search.
39	Prompt Cache Stability	Shows how stable prompt prefixes, byte-equal forked requests, and execution-time permission checks improve Agent prompt-cache reuse.
40	Durable Agent Loops	Designs resumable Agent turns with phase-boundary checkpoints, interrupted markers, bounded unattended recovery, and idempotent persistence.
41	Steering a Running Agent	Explains how running Agents absorb follow-ups, preserve message addressing, distinguish delivery acknowledgement from replay, and interrupt safely.
42	Agent Timeouts and Watchdogs	Separates soft-idle warnings, hard cancellation, stream-gap detection, and phase-aware watchdog policy for hung Agents.
43	Stuck Loop Detection	Builds stuck-loop detection from complementary progress signals, bounded nudges, force-stop behavior, and a public zero-byte write incident.
44	Parallel Tool Execution	Explains safe parallel tool calls using per-invocation concurrency traits, ordered result pairing, and conservative Bash classification.
45	Computer-Use Context Management	Separates text observations, browser screenshots, and global image budgets so GUI loops retain actionable state without unbounded prompt growth.

Reading paths

Context and cost: 35 → 36 → 37 → 38 → 39
Durability and control: 40 → 41 → 42
Reliability, speed, and GUI context: 43 → 44 → 45
Coming from DAG workflows: 34 → 40 → 43 → 44

Source boundary

Only tracked files on Kocoro public origin/main are used. Local-only architecture notes, unpushed branches, private postmortems, hosted-service topology, internal tokens, and private repository paths are excluded.

Chapter 45 is included because its observation-window and image-budget implementation is now tracked on public origin/main.

@incollection{zhang2026aiagent_en_part10-agent-loop-engineering, author = {Zhang, Wayland}, title = {Part 10: Agent Loop Engineering}, booktitle = {AI Agent Architecture: From Single Agent to Enterprise Multi-Agent Systems}, year = {2026}, url = {https://waylandz.com/ai-agent-book-en/part10-agent-loop-engineering/} }