Lecture 06. Agent topology, context and memory as recorded decisions

Part 2 · The Agent as Tested Subsystems · Scheduled 13 Dec 2026 · 1-hour lecture

Upcoming course outline. This class has not been recorded as studied.

Learning objectives

Topology catalogue: single loop, router, orchestrator-worker, evaluator-optimizer, plan-then-execute; three reference architectures (compound AI, SWE-agent, brain / hands / session).
Static vs dynamic context (Day 1 pp. 15-17); compaction vs observation masking; context rot.
Memory types; update actions (ADD, UPDATE, DELETE, NOOP); memory poisoning (OWASP ASI06); a progress file a second session resumes from.
The ADR: name the quality attribute traded, cite one measurement, add one C4 container diagram with the trust boundaries from classes 4 and 5.

Industry: Long tasks fail on context, not on the model; students choose a policy with a measurement, not a guess.
Academic: Masking vs compaction is an open research question (The Complexity Trap); the lab arms are a direct replication.
Elsewhere: Students understand what /compact and memory files do inside a coding agent.

Stanford CS329Z L6 and L7 slides.
Google Day 1 2026 pp. 15-17; Google Context Engineering: Sessions & Memory 2025.
The paper (Lindenbauer et al. 2025) as pre-reading.

These are preparation requirements in the draft; staff materials are not bundled with this site.

The linked reference library contains the source URLs extracted from the course PDF. Availability notes above reflect the draft of 4 November 2026.

Continue with Lab 06: Agent topology, context and memory as recorded decisions. Every lab contributes an artifact to the same repository and a measurement to the results ledger.

[Lecture notes and observations]