Vibeleaderboard
← Course overviewModule 09 · Coordinate several pieces of work · Beta
Lesson 9.5

Module contents

Coordinate several pieces of work

Lesson 5 of 5

Is splitting the work worth it?

Four assignments, four evidence-bearing returns, a reconciliation, an integration step with checks. Each of those took time and, where model use is paid for, money, and the treasurer could have done the supplier decision in one session. This lesson counts what coordination costs against what it speeds up, looks at what practitioners report about the costs of running several agents, and ends with the coordination plan that records the decision, a transfer task in a different setting, and the bridge to the final module.

Numbered links identify sources. Underlined terms offer a definition without leaving the lesson.

What coordination costs

The of splitting work is what the split adds that a single worker would not have done: writing the assignments; the workers' repeated reading of the shared inputs, since each loads the rubric and constraints into its own context; the coordinator's checking of the returns against each other, including the deliberate cross-check on the two overlapping fields; the waiting at the fan-in, since the final text cannot be completed until the slowest return is in; and the merging in the integration step. Checks that the work needed anyway, such as tracing every figure to a passage, are not coordination cost; a single worker's report would have needed them too. Only the duplicated and added effort counts.

Against that is the time saved, and the map bounds it. Only the parts that ran at once were accelerated. Here is the treasurer's estimate, authored to show the arithmetic, with durations invented:

Part
Write assignments
One worker, in sequence
0
Four workers, in parallel
15 min
Part
Facts, three suppliers
One worker, in sequence
3 × 20 = 60 min
Four workers, in parallel
20 min (longest of three, at once)
Part
Eligibility
One worker, in sequence
15 min
Four workers, in parallel
at the same time as the facts
Part
Check returns, reconcile
One worker, in sequence
10 min
Four workers, in parallel
25 min (four returns, origins compared, overlap cross-checked)
Part
Recommendation and checks
One worker, in sequence
25 min
Four workers, in parallel
25 min
Part
Total elapsed
One worker, in sequence
110 min
Four workers, in parallel
85 min

The split saved 25 minutes, a little under a quarter of the elapsed time, because the gathering and the eligibility check ran at once and the rest of the map is a line. It also spent more total effort, since four workers each read the inputs and the coordinator did more checking, which matters when effort is paid for. Change the numbers and the answer changes: if each supplier's facts took an hour, the split would save most of the afternoon; if the recommendation took most of the time, it would save almost nothing. The decision is arithmetic on the map, not a rule about agents.

Costs practitioners report, and their scope

Two vendor accounts give figures, and each belongs to its setting. Anthropic's June 2025 research-system account reports token use on the order of fifteen times an ordinary chat's and says the arrangement pays only on tasks valuable enough to bear it; its February 2026 compiler account reports roughly two thousand agent sessions and about twenty thousand dollars of model use. 2 3 Neither is a rate a reader can apply elsewhere. What they establish is that parallel agent work costs tokens and money as well as time, and that a coordinator who counts only elapsed time is missing a column.

A third source is about organizing rather than cost. A post on X in August 2026 by Pieter Levels, a developer who runs several products alone, describes keeping one terminal multiplexer session per project on a remote server and attaching to whichever project he is working on from his phone. 6 The site's Intel entry reproduces the post; the original could not be fetched when this module was checked. It is a dated personal preference for keeping projects' sessions apart so that their contexts do not mix. It is not, by itself, folder or permission isolation, which the previous lesson showed a session does not provide, and it is not evidence about what survives a disconnection or a failure, which the next module takes up with the sources that address it.

The coordination plan

The record this module produces is a , which sits beside module 8's workflow review pack and refers to it. Here is the treasurer's, in outline:

Coordination plan: spring supplier decision, October 2026

Dependency map:        as drawn in lesson 1 (four parallel parts, then a line;
                       outline of the recommendation drafted early)
Decision:              split the four parallel parts; one coordinator (treasurer's
                       session) owns reconciliation and integration
Assignments:           F-Greenway, F-Marlow, F-Fernbank, E-all, R (lesson 2)
Shared inputs:         suppliers.txt, rubric.txt, constraints.txt in the shared folder
Deliberate overlap:    delivery and invoice fields read by both F-* and E-all,
                       as a cross-check; cost counted below
Handoffs:              each return = output file + saved pages in the worker's own
                       folder + passages + uncertainty
Isolation:             each worker writes only its own output file and page folder;
                       recommendation.txt written once, by R; workers fetch the assigned public pages but make no changes
                       to outside services
Integration checks:    every figure traceable to a passage or a shown derivation
                       (program); no two values from one origin counted twice
                       (coordinator); overlap fields compared and settled at the
                       page (coordinator); recommendation follows from the table
                       (person); Fernbank invoice terms unknown until confirmed
Cost estimate:         85 min elapsed vs 110 single; inputs read four times;
                       paid model use per worker if metered
Stop conditions:       any worker's stop (page dead, no prices) halts the join
                       until the treasurer decides; a return without origins is
                       sent back
Review after the run:  did the split save what the estimate said? record actual
                       times and what the cross-check caught

The plan is checked after the run against what happened, the way module 8's comparison checked a proposed improvement. If the split saved less than estimated, or the integration found disagreements that cost more than the parallel gathering saved, the next decision uses the actual figures.

Transfer checkpoint: a grant application

Try the module's distinctions on a different piece of work. Everything here is authored.

The garden is applying for a council grant. The application has three parts: a budget, which needs written quotes for two items of equipment; a narrative describing the project; and an appendix of supporting evidence, including letters of support and any figures about local need. The narrative has to cite the budget total and one figure from the appendix. Three workers are proposed, one per part.

When the returns arrive, the appendix worker has two letters of support that both cite the same council survey of local green space, one from a school and one from a residents' association, and reports them as two sources for the need figure. The budget worker returns "Budget complete: £4,380, both quotes obtained" and nothing else. The narrative worker returns a draft that cites a budget total of £4,200 and gives the survey's need figure to one decimal place.

Before reading the guide, answer five questions. First, draw the dependency map: which parts can run at once, and which parts of the narrative depend on the others? Second, how many independent observations of the need figure are there, and what would corroborate it? Third, what is missing from the budget worker's handoff, and what can the coordinator do with it as it stands? Fourth, the narrative and the budget disagree on the total; how is that resolved, and what does the disagreement show? Fifth, could a fourth worker help, and what must the coordinator review centrally before the application is sent?

Self-review guide. On the map: the budget and the appendix are independent and can run at once. The narrative's structure and most of its text depend on neither, and can be drafted early; its two cited figures depend on the budget and the appendix and cannot be final until those are in. A complete answer draws that distinction rather than putting the whole narrative after the others.

On the need figure: there is one independent observation, the council survey, reached through two letters. The letters remain valuable as support, which is a different role from evidence for a figure. Corroboration would need a second origin with its own access to the fact: independently collected comparable data, such as a count the garden made or a different body's study. A page of council statistics may be the same survey published again, and has to be checked for that before it counts.

On the budget handoff: it is a conclusion without rows. No quotes named, no passages, no uncertainty, no artifacts, so the coordinator can only accept, redo or ask. The missing handoff is the two quotes saved in the folder, each with the supplier, date, amount and what it covers, and a line for anything the quotes leave out.

On the disagreement: the budget's quotes are the origins, so the total is resolved by opening them and adding, not by choosing between the two workers' sentences. The disagreement shows that a figure in the narrative was not traced to its origin; it does not by itself show that the narrative started too early, since a narrative started late could have copied the same wrong figure, and the cure is the trace, not the timing.

On a fourth worker: the answer depends on finding independent work for it. Gathering the two equipment quotes from two suppliers is independent work that could be split, and reviewing the finished draft against the criteria gives a fresh reader a chance to catch mistakes the writer missed. A fourth worker on the narrative's final figures would wait. The central review, whoever does it: every figure in the narrative traceable to the budget or the appendix, the need figure counted once, the quotes opened and the total recomputed from them, and the narrative read by a person for whether it follows. An answer that adds a worker without naming the independent work it would take, and without counting the coordination it adds, is not yet complete.

What comes next

The supplier decision ran on the treasurer's laptop, in an afternoon, with the treasurer watching. Coordinated work is often not like that. A coordinator that waits for four returns may wait hours; a worker that fetches fifty pages may still be running when the person who started it closes the laptop and goes home. The work then continues somewhere, or stops, and the person supervising it from a phone is looking at a window onto a process on another machine. Where that process runs, what it can reach, what happens when the connection drops and what happens when the machine fails are the final module's questions, and every distinction in this module, the map, the handoffs, the isolation, is carried into it as it is.

Carry this forward

  • Coordination cost is the effort a split adds: assignments, repeated input reading, cross-checking returns, waiting at the fan-in, merging. Checks the work needed anyway are not overhead. Only the parts that ran at once were accelerated, so the saving is arithmetic on the map, and total effort usually rises even when elapsed time falls.
  • Vendor accounts report parallel agent work costing many times a single session's tokens and, for one project, thousands of sessions; those are their figures for their tasks and establish only that the costs are real.
  • One session per project is an organizing preference that keeps contexts apart. It is not by itself folder or permission isolation, and it says nothing about surviving a disconnection or a failure.
  • A coordination plan records the map, the decision, the assignments, the deliberate overlaps, the handoff form, the isolation rule, the integration checks, the cost estimate and the stop conditions, and is checked against what the run actually cost.

Lesson check

A committee report has four sections that must each be written after the previous one is agreed, and a short appendix that can be compiled at any time. Someone proposes five workers, one per section and one for the appendix, to finish sooner. What does this module say?

Choose an answer for each question.

This is a practice check. Answers reset when you leave this page. Results are not saved and do not place you in another module or award a qualification.

You have reached the end of module 9. The module quiz below checks how these ideas fit together before module 10 asks where the work actually runs when you are not there.

Module 9 quiz: put the ideas together

1. The treasurer is asked "how many agents should we use for the supplier decision?" What does this module say to do first?
2. In the framework example, two parallel branches each tried to write the same piece of shared state in one step and the framework refused. What is the lesson for a document workflow?
3. Why does assignment F-Greenway tell the worker to look at Greenway only?
4. Two returns agree that Greenway's delivery charge is £18.00. One quotes Greenway's price page; the other quotes a newsletter that reprinted Greenway's announcement. What has the agreement established?
5. A worker returns "Marlow is not eligible." and nothing else. What can the coordinator do with it, and what is lost?
6. Two workers in separate chat sessions both have a tool that rewrites recommendation.txt as a whole file. Does running them in separate sessions protect the file from an edit collision?
7. A worker is named "reviewer" in its assignment. What limits what it can actually do?
8. Splitting the supplier decision across four workers saved 25 minutes of 110 in the authored estimate, a little under a quarter. Why not more?

Choose an answer for each question.

This is a practice check. Answers reset when you leave this page. Results are not saved and do not place you in another module or award a qualification.

Sources · Reviewed 2026-09-29
  1. 2. Anthropic engineering (13 June 2025): How we built our multi-agent research system
  2. 3. Nicholas Carlini, Anthropic (5 February 2026): Building a C compiler with a team of parallel Claudes
  3. 6. VibeLeaderboard Intel: Pieter Levels on X (24 August 2026), One tmux session per project: running coding agents from a phone

Sources support nearby explanations. Product details can change. Scenarios and model results are authored illustrations unless a lesson explicitly identifies a recorded execution. Module 7 identifies its offline code checks.