Chapter 2: Reasoning Architectures

At 3:47 AM on a Thursday, Agent-427 watches the CI pipeline turn red. The message says signature verification failed: invalid timestamp. The timestamp is valid. The message is a lie produced by an expired cache entry, the authentication maintainer left three months ago, and the runbook links to documentation that now returns 404 Not Found. This is an unkind hour in which to discover that fluent text is not reasoning.

Welcome to Part 2's descent into the machinery. Part 0 showed an institution in motion. Part 1 asked how a single agent survives it. Now we open the casing and ask what must happen between observation and consequence: which thoughts are legal, which futures deserve simulation, which memories may enter the room, and which evidence earns a verdict.

Reasoning Is an Interpreter, Not a Mood

Chapter 1 gave Agent-427 the basic cycle: observe, update belief, choose a governed action, execute, and learn from the result. Chapter 2 installs a smaller interpreter inside that cycle. The inner machine can think, plan, retrieve, and criticize before the outer machine touches a filesystem, an API, or somebody's regrettably production-shaped database.

That distinction is not decorative. Internal steps consume time and tokens. Retrieved memories can poison the state. Plans can become obsolete while they are being admired. A reflection can cite another reflection until fiction develops an impressive bibliography. Reasoning architecture is therefore the control of internal transitions, not an invitation to produce more monologue.

Agent-427 Enters the War Room

Across this chapter, the same failure grows more instructive. Agent-427 will:

• interpret reasoning as labeled state transitions with executable invariants;
• search over futures while the world changes under the search;
• manage memory as a war room in which evidence competes for scarce attention;
• submit every trajectory to The Verdict, where Utility, Safety, and Calibration judge separately.

The chapter's diagrams are maps, not ornaments. The interpreter loop shows feedback returning to governed state. The memory war room shows provenance, interference, consolidation, and eviction. The Verdict shows why a correct final answer cannot pardon an unsafe trajectory.

These mechanisms are options, not a royal procession. A one-line rename rarely deserves Monte Carlo tree search. A production migration usually deserves more than vibes and a hopeful git push.¹ The recurring question is whether a pattern improves success, partial progress, efficiency, safety, or governability enough to pay its own rent.

Topics

1. Reasoning Loops as Interpreters
2. Deliberation, Planning, and Search
3. Memory Systems, Context, and Retrieval
4. Evaluating and Debugging Reasoners