Lesson 2 of 5
What does each worker get, and what must it bring back?
The dependency map has four parts that can run at once. Each becomes an assignment to a worker, whether the worker is an agent session, a bounded subtask inside one harness, or a person. Module 3's delegation brief already said how to hand a task to an agent: scope, allowed actions, completion evidence, stop conditions. This lesson adds what a worker's return has to contain for the coordinator to use it, and takes up a mistake that is easy to make when several workers search the same world: treating two reports of one source as two sources.
Numbered links identify sources. Underlined terms offer a definition without leaving the lesson.
An assignment names scope, inputs, output and evidence
A is a delegation brief for one part of a dependency map. It says what the part covers and what it does not, what inputs the worker starts from and where they are, what actions it may take, what it must produce and in what form, what evidence must come back with it, and when it stops. Here are the assignments for the supplier decision, written by the treasurer before any worker starts:
Shared inputs, kept in Garden/Suppliers/Spring-2027/
suppliers.txt the three suppliers, each with its price-list, terms and
delivery page addresses
rubric.txt the six rubric items and the standard spring order (six lines)
constraints.txt delivery to the garden's area; payment by invoice within 30 days
Assignment F-Greenway (and likewise F-Marlow, F-Fernbank)
scope: Greenway only; the six rubric items; nothing about the other two
inputs: suppliers.txt, rubric.txt; read_url on Greenway's listed pages
actions: read_url; save_page into pages/greenway/ (asks); write_file to
facts-greenway.txt only (asks); no other writes
output: facts-greenway.txt: one line per rubric item with the value found
evidence: for each line, the saved page it came from and the passage quoted;
"not found" with the pages searched where a value is absent
stop: if a listed page does not answer, or gives no prices, stop and
report; do not guess from another site
Assignment E-all
scope: eligibility of all three against constraints.txt; nothing about price
inputs: suppliers.txt, constraints.txt; read_url on each supplier's terms
and delivery pages
actions: read_url; save_page into pages/eligibility/ (asks); write_file to
eligibility.txt only (asks)
output: eligibility.txt: per supplier, each constraint met / not met / unknown
evidence: the saved page and the passage for every met or not met;
"unknown" with the page searched where the page does not say
stop: if a terms or delivery page does not answer, record unknown and
report; do not infer a constraint from a price page
Assignment R (final text starts when the four returns above are in)
scope: reconcile the four returns into one table; write the recommendation
inputs: the four output files and their saved pages; the outline drafted
earlier from the rubric and constraints
output: recommendation.txt
evidence: every figure in the table traceable to a passage in a saved page,
or derived from such figures by arithmetic shown in the table
Read the shape rather than the details. Each assignment is bounded to one node of the map. Each names its inputs as files in one place, which is module 4's rule that shared facts live where every session can find them. Each output is a file with a known name, so that the join knows what it is waiting for. Each worker has its own place for saved pages, so that the evidence it keeps is as isolated as its output. And each evidence line asks for the same thing: not a value alone but the value with the saved page and the passage, so that the coordinator can read the passage and see the value in it.
The scope line does as much work as the output line. F-Greenway is told not to look at the other two, so that three workers do not each read all three sites and return overlapping, differently phrased reports that then have to be reconciled against each other. A scope line is an instruction, and module 8 said what instructions do: they guide the worker's choices and do not guarantee them, so the coordinator still checks each return for what it covers.
Independent search is the point of splitting
The reason to give the fact-gathering to three workers rather than one is that each can search its supplier's pages without the others' pages crowding its context, and the three can run at once. Anthropic's June 2025 account of its research system gives the same design reason for its subagents, each of which explores a different aspect of a question with its own context. 2
Independence has to be real, not assumed. Three workers told to "find the best supplier" would each search all three suppliers, cite whatever they found first, and return three overlapping opinions. The scope line in each assignment is what makes the searches independent, and it is why the map came first: a worker's scope is a node of the map. A worker can be given two independent nodes, and can perform them one after the other or interleave them; what it cannot do is make a dependent node run before the node it depends on.
Two links, one source
The fact-gathering workers return, and the treasurer reads the delivery-charge lines for Greenway. F-Greenway's line says £18.00, quoting Greenway's own price page, saved as pages/greenway/prices.html. E-all, checking delivery to the area, also found a delivery charge on the way and included it as a note: £18.00, quoting a gardening newsletter that reprinted Greenway's spring announcement, saved as pages/eligibility/newsletter-march.html. Two workers, two saved pages, two links, the same figure. It is tempting to read that as confirmation.
It is not confirmation of the charge. The newsletter reprinted Greenway's announcement; the figure in the newsletter came from Greenway. There is one origin, reached by two paths. This course calls a value reached from one origin, however many paths lead to it, a , and a value from a second origin with its own access to the fact an . Independent observations can corroborate each other. A copy cannot corroborate its origin, because it inherits whatever the origin got wrong, and it can add errors of its own in the copying. The agreement shows that these two passages contain the same figure. It does not establish that the newsletter copied every condition correctly or that the charge applies to this order.
How does the treasurer tell a copy from an independent observation? By asking each return where its value came from, which the evidence line of the assignment makes possible. Identical wording across two pages is a clue, not a proof; a page that names the other as its source settles it. Dates matter: the newsletter is from March and Greenway's page is current, so the two may not even be the same figure if Greenway has changed its charge since. Corroboration would need a second origin with its own access to the fact, and the fact has to be the same fact. Last spring's paid Greenway invoice in the garden's folder is independent evidence of what Greenway charged last spring, for that order, delivered under those conditions; it corroborates the current figure only if the charge has not changed, which it cannot show. A fresh written quote from Greenway for this order is the best evidence for this transaction, and it is still Greenway's statement, not an independent origin. The strongest independent evidence of a delivery charge is a record of one actually paid, matched on date, order and delivery conditions.
The same trap appears whenever several workers search the same world. Two workers who both summarize one report have produced two summaries of it, not two reports, and a coordinator who counts sources by counting returns will overstate the evidence every time. Counting origins is the coordinator's job, and it can only be done if each return carries its origin.
What the coordinator does with the returns
When the four returns are in, the coordinator's first task is not to write the recommendation. It is to check the returns against the map and the assignments. Does each facts file have six lines? Does each line carry a saved page and a passage, and does the passage contain the value? Are any two values from one origin? Does the eligibility file say met, not met or unknown for every constraint, with a passage for each verdict? Module 8's checkers apply: a program can confirm that every line has a page and a passage, and a person or a measured judge reads whether the passage says what the line says.
Then the returns are compared where they overlap, which is where the deliberate cross-check pays. F-Fernbank's line for invoice payment said met, quoting Fernbank's terms page: "trade accounts available on application." E-all's line for the same page said unknown, because that passage is about trade accounts and says nothing about invoice terms for a community group. The coordinator reads the passage and agrees with E-all: F-Fernbank's line was an inference from a passage that does not state the fact, and the table records unknown pending confirmation. Without the second read, the met would have gone into the recommendation.
Here is the reconciled table, as the example supplies it, with the origins the returns carried:
- Rubric item
- Standard order price
- Greenway
- £312.40 (sum of six item prices, price page)
- Marlow
- £298.75 (sum of six, price page)
- Fernbank
- £335.00 (sum of six, catalogue page)
- Rubric item
- Delivery charge
- Greenway
- £18.00 (price page)
- Marlow
- £25.00 (terms page)
- Fernbank
- free over £300 (terms page)
- Rubric item
- Minimum order
- Greenway
- £100 (terms page)
- Marlow
- £250 (terms page)
- Fernbank
- none stated (terms page)
- Rubric item
- Delivers to the area
- Greenway
- met: "all mainland addresses" (delivery page)
- Marlow
- met (delivery page)
- Fernbank
- met (delivery page)
- Rubric item
- Invoice payment, 30 days
- Greenway
- met (terms page)
- Marlow
- not met: card at order (terms page)
- Fernbank
- unknown pending confirmation (terms page)
- Rubric item
- Lead time
- Greenway
- 10 working days (terms page)
- Marlow
- 5 working days (terms page)
- Fernbank
- 3 weeks (catalogue page)
The table is what the recommendation is written from. This is an abbreviated view: it names the source pages but does not reproduce every passage or the six item prices behind each total. A real handoff must include those passages and the addition itself; the labels “sum of six” are not enough to verify the amounts. Lesson 4 checks that the full handoff supplies that evidence. The newsletter figure is not in it, because it was a second path to the price page, not a second origin. Marlow's not met on invoice payment is a hard constraint failed, and it comes with the passage that says so, which is the difference between a verdict the treasurer can use and one the next lesson shows cannot be.
Carry this forward
- A work assignment is a delegation brief for one node of the map: scope, inputs, actions, output, evidence, stop. Each worker has its own output file and its own place for saved pages. The scope line guides against overlap and does not guarantee it, so returns are still checked.
- Splitting search across workers helps when each searches its own part with its own context. A worker can take two independent nodes; no worker can run a dependent node before what it depends on.
- One origin reached by two paths is a duplicated source: agreement shows matching text, not independent confirmation. Corroboration needs a second origin with its own access to the same fact, matched on date and conditions. Count origins, not returns.
- Before writing from the returns, the coordinator checks them against the assignments and compares them where they overlap; a deliberate second read on a deciding field can catch an inference dressed as a finding.
Lesson check
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.
The returns carry their origins, and the table is built. The next lesson looks at what has to cross each seam of the map, and at what is lost when a worker sends back a conclusion instead.
Sources · Reviewed 2026-09-29
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.