AI Assisted Software Engineering

Part II · The Life Cycle, Audited

08 Coordination at Scale: Shared Context vs. Shared Messages

HoldsLab: Coordination designFeeds M3

Load-bearing assumption

Coordination happens through bandwidth-limited messages

Verdict: Holds

Coordination theory assumes workers coordinate through bandwidth-limited messages; agents can share context instead, but shared context carries only what was fixed before the work began, so explicit channels remain the central engineering problem.

Learning objectives

After this chapter you should be able to:

  1. Decompose a small system by hidden design decision rather than by processing step, and show with Parnas's KWIC example why the first approach reduces what workers must say to each other.
  2. State Conway's hypothesis and summarize the evidence for it and the evidence that qualifies it.
  3. State the load-bearing assumption of coordination theory and apply the negation test to it.
  4. Distinguish shared context from shared messages, and describe four agent architectures by who communicates with whom.
  5. Produce a coordination design for a small agent team: a decomposition, a context-or-message decision for each interface, an escalation matrix and a permission set.
  6. State what evidence would overturn the verdict.

Opening case

A composite case, not a real engagement.

A consultant and two engineers are building a dispatch dashboard for a regional logistics company. They split the build across three coding agents that run in parallel. All three use the same large language model (LLM), read the same specification and work in the same repository under the same convention file. "They share everything," the consultant says. "They can't disagree."

The chapter continues from here.

What the chapter covers

  1. 8.1 The classical method
    • 8.1.1 Parnas: decomposition as a coordination device
    • 8.1.2 Conway: the organization becomes the system
    • 8.1.3 Brooks: communication overhead, and the logged channel
    • 8.1.4 What the theory got right
  2. 8.2 The assumption it depends on
  3. 8.3 What agents change
    • 8.3.1 Shared context as an apparent alternative
    • 8.3.2 Four architectures
    • 8.3.3 The architecture debate
    • 8.3.4 Handoffs and the missing overhearing
    • 8.3.5 What shared context cannot buy: a preview of the bound
  4. 8.4 The verdict
  5. 8.5 What replaces it: a coordination design method
    • 8.5.1 Dispatch, redesigned
  6. 8.6 Lab: Coordination design
  7. 8.7 What would make this verdict false

Lab and prediction

Lab: Coordination design

Feeds milestone M3 (the deliverable). How the labs fit the project

This chapter's own prediction: Most coordination defects trace to undeclared dependencies, not to failures of declared channels.

A result that contradicts a prediction, honestly reported, is a result. You are graded on accurate records and reasoned interpretation, not on agreeing with the book.

Summary

  • Decomposing by hidden decision, not processing step, confines changes to one module: by my tally of Parnas's example, five likely changes touch at least 16 modules one way and five the other. Conway's mirroring is a qualified tendency, and Brooks's remedy is explicit, logged channels.
  • Load-bearing assumption: coordination happens through bandwidth-limited messages. Shared context carries what was fixed before the work; information generated during the work needs a message, and the bandwidth limit survives in tokens, context and review attention.
  • Verdict: holds, and coordination becomes the central engineering problem for agent teams. Coordination with no visible channel is hidden coupling.
  • A handoff packet is a message with a structured body; its context summary behaves like compressed shared context and is a claim to check.
  • The CHSH bound is a preview, not support for the verdict: its software mapping is a conjecture (Chapter 17).
  • The extension is a seven-step design method, from listing decisions to setting permissions.

Key terms

Information hiding
designing each module to conceal one design decision from the rest of the system (Parnas 1972).Also used in 1, 2, 10
Hidden coupling
a dependence between workers' outputs that no declared interface and no logged channel carries.Also used in 1, 16, 17, 18
Shared context
everything agents hold in common before a task starts.Also used in 1, 2, 16, 17, 18
Shared message
information one agent produces during the work and sends to another.
Single agent with full context
one agent holds the whole task; no inter-agent channels.
Orchestrator–worker
a lead agent decomposes, delegates and integrates; workers talk only to the lead.
Message passing
agents exchange explicit messages in agreed formats; the pipeline is the common case.
Blackboard
agents read and write a shared store while a control component decides who acts.
Handoff packet
a structured bundle of work artifact, evidence bundle, context summary, task specification and handoff metadata (Hassan 2026, §6.6.4, p. 201).Also used in 16, 17, 18
Congruence
the match between the dependencies that exist and the channels that carry them.
Escalation matrix
decision type, owner, consult path and required evidence.Also used in 16, 18

Connections

Where this chapter and another point to each other most often.

See it on the map

Further reading

  • Parnas (1972). Ten pages. Read the KWIC comparison until you can restate why the second decomposition is cheaper to change.
  • Conway (1968). Four pages. Read the part on why the first design decision is the organization.
  • Herbsleb and Grinter (1999). What happens when informal channels break; the human-team analogue of the missing overhearing in Section 8.3.4.
  • Colfer and Baldwin (2016). A balanced review of the mirroring hypothesis, including when it does not hold.
  • Anthropic (2025) and Yan (2025), paired. The two sides of the architecture debate. Ask what each measured.
  • Cemri et al. (2025). The failure taxonomy for multi-agent systems; read the inter-agent misalignment section.