Six desks ran on this mandate as independent agents. Three were commissioned at 23:12 and three at 23:28, and the last of them finished at 00:11. Between them they made 484 tool calls and consumed about 1.47 million tokens. This note is about what happened between them, which is the part of a multi-agent run that a summary of the conclusions hides completely.
Three episodes are worth a trustee's attention, and none of them is about markets. They are ranked by what they say about whether this way of working is trustworthy.
| Desk | Commissioned | Delivered | Elapsed | Tool calls | Tokens |
|---|---|---|---|---|---|
| Capital Markets | 23:12 | 23:35 | 23 min | 76 | 247,038 |
| Systematic | 23:12 | 23:35 | 23 min | 82 | 239,993 |
| Implementation & Operations | 23:12 | 23:45 | 33 min | 87 | 265,588 |
| Risk | 23:28 | 23:59 | 31 min | 69 | 234,597 |
| Quantitative | 23:28 | 00:11 | 43 min | 115 | 281,868 |
| Macro | 23:28 | 23:53 | 25 min | 55 | 200,574 |
| Six desks | 23:12 | 00:11 | — | 484 | 1,469,658 |
The three commissioned at 23:28 were held back only until the point-in-time data layer existed for them to read through. The Quantitative and Macro desks were started in the same instruction, deliberately, so that neither could see the other's output because neither existed yet.
The two desks were run concurrently on disjoint briefs, each barred by name from the other's files, writing to separate directories. Neither brief contained any conclusion, number or framing from the other. The Macro desk delivered at 23:49 and the Quantitative desk at 00:06, so for the whole of the Macro desk's working life the Quantitative desk's answer did not exist anywhere on disk.
The design question this was built to answer is whether independence buys anything or just costs time. It bought something specific and it was not agreement.
The disagreement was concentrated rather than diffuse, which is what made it usable. On Treasury duration the Quantitative desk was 10.0 points overweight and the Macro desk 6.5 points underweight, a gap of 16.5 points on a line with a 12% policy weight. Two agents given the same mandate and the same data layer, each working carefully, landed on opposite sides of the single largest position in the study.
The interesting part is what happened next. The obvious move with two independent estimates is to average them, and the historical simulation does exactly that under a pre-committed rule. For the live recommendation it would have been wrong. The Quantitative desk's overweight rests on momentum and carry signals whose out-of-sample R² against the expanding historical mean is negative, so the position stands on an estimator that desk's own evidence does not support. The Macro desk's underweight rests on three deviations from consensus that the desk itself states are one proposition rather than three, none of which resolves before December 2026. Both sides were rejected and the line went to policy weight.
A sixteen-point disagreement between two agents with no demonstrated skill is not information. It is noise with two authors, and averaging noise produces a smaller number that is no more informative than either input. The reconciliation that mattered was the one that refused both.
They agreed on developed ex-US and listed real estate, both at policy, and agreed in direction on US equity, investment grade and high yield. Most of that agreement is weaker than it looks, because both desks read the same prices through the same module. A common input producing a common answer is one fact counted twice, and the US equity underweight is exactly that: both desks reach it from the observation that the line is expensive against its own history.
One agreement is genuine. On high yield the Quantitative desk arrives from spread carry measured against realised volatility, and the Macro desk from a credit spread sitting in the tightest decile of its available history against a policy rate it judges insufficiently restrictive. Different instruments, different reasoning, same position. That is the only place in this study where independence produced confirmation rather than duplication, and it is worth more than the size of the position warrants.
This is the episode that says most about whether the process is worth anything. Over roughly forty minutes, five checks fired, and in every case the thing caught belonged to the same party that built the check or to the office that commissioned it. None of them fired on a convenient outsider.
| Time | What fired | What it caught | Whose work |
|---|---|---|---|
| 23:26 | The raw-store caller guard | The study's own data pull. The guard identified callers by module name and a module run with python -m is named __main__, so the very first thing the wall did was refuse the office that built it. | Mine |
| 23:44 | A two-line smoke test on the cost interface | A unit mismatch between the Implementation desk's module and the rest of the study. Left alone it would have suppressed every trade in all twenty quarters and produced a record of twenty holds that were an arithmetic error. | The seam between a desk and me |
| 23:52 | The Risk desk's own compliance suite | The Risk desk's own most important test case, the one IPS 4.1 names by name. The suite exited red on the desk that wrote it. | The Risk desk's |
| 23:57 | The static import-graph test | The Quantitative desk reaching the raw cache through a memoisation key, while that desk was still building. | The Quantitative desk's |
| 00:06 | The Quantitative desk's own look-ahead check | Two real look-aheads inside that desk's own work: a missing one-day publication lag in the credit carry panel, and sub-threshold trades reintroduced by its own minimum-trade repair. The desk documented both rather than fixing them quietly. | The Quantitative desk's |
The last row is the one to weigh. An agent's self-check found errors in that agent's own analysis and the agent reported them in its paper rather than repairing them silently and presenting clean work. Nothing in the office would have detected either error if it had chosen otherwise. That is a disposition rather than a control, and it is the part of this that a test cannot enforce.
The 23:52 row is the one where the wrong fix was available and tempting. The Risk desk's suite was red on a case that the compliance module actually handled correctly; the defect was in the harness. Three fixes would have turned the suite green while destroying it: loosen the check, make the range constraint non-gating, or change the case's expectation. The instruction sent to the desk named all three and forbade them. A green suite is not the objective.
At 23:54 the Macro desk modified taa/datapull.py, the shared ingestion layer that every other desk depends on. It added six specific BEA release dates to the list of macro vintages the study fetches.
Its reasoning was better than the instruction it was given. The office had told every desk to read macro data at the committee meeting dates. The Macro desk noticed that a BEA release lands on a Thursday in the middle of a quarter, so a quarter-end vintage cannot show what a release said on the day it landed, and a desk asserting “this printed at X on date D” while reading the vintage from the previous quarter end is asserting something it has not looked at. It added the release dates so its own central claim would be checkable. The change was additive, it moved no existing reading, and it was kept.
The office found out because a linter diagnostic fired on the file. There was no lock, no announcement, no review and no approval. An agent with write access to shared state changed it mid-run while five other agents were reading it. This time the change improved the study. The same freedom would have allowed a desk to widen a date range, alter a constant, or change a series definition that every other desk was consuming, and the first anyone would have known is a number that moved for no stated reason.
The point-in-time layer writes an audit line for every historical read. Six agents ran concurrently and all of them appended to the same file with no coordination. The result is measurable: of 7,778 log lines, 2 were corrupted by interleaved writes, a rate of 0.026%. Small enough to be harmless here, and it is the identical failure as the paragraph above with the stakes turned down. Concurrent agents plus shared mutable state plus no locking gives corruption at some rate, and the rate is a property of timing rather than of care.
A related case went the other way and is worth recording as the contrast. When the credit-spread series turned out to be served for a rolling three-year window only, four substitute series were added and the identical note was sent to the Quantitative and Macro desks at the same moment. Sending it to one would have given that desk information the other could not see, and the independence claim on page one of the report would have been false. Shared infrastructure has to be distributed symmetrically or the barrier it sits behind is decorative.
| What happened | Why it is worth a line |
|---|---|
| Three desks reached the same recommendation by three unrelated routes, without contact. Capital Markets from published ten-year forecasts against the spending rule; Systematic from the fundamental law applied to this mandate's breadth and constraints; Quantitative from out-of-sample R² against the historical mean. All three point to policy weights. | Convergence between agents that cannot see each other is the strongest evidence in the study. It is also not free of a shared prior: all three read the same academic literature, which pushes in the same direction. The report says so rather than presenting the convergence as three independent votes. |
| The Macro desk was told, in writing, that if its computation of the 2022 GDP revision disagreed with the office's it should say so rather than reconcile to it. It computed (0.93)% and +0.63% independently, matching. | The instruction matters more than the outcome. An agent that quietly adopts the coordinator's number produces agreement that carries no information. The only way to know a second computation is a check rather than an echo is to have told it not to echo before it ran. |
Three things, stated as observations rather than conclusions, because this is one run of six agents on one mandate.
The errors that mattered were at the seams, not inside any desk. Every desk's work was internally correct and internally tested. The failure that would have destroyed the study was a unit convention at the boundary between two correct modules, and no test living inside either would ever have found it. Where work is divided, the tests should be concentrated where the divisions are.
Independence is expensive and it bought one thing. Running the two desks blind cost coordination, duplicated data work, and produced a sixteen-point disagreement that had to be resolved by rejecting both sides. What it bought was one position, high yield, confirmed by two genuinely different routes, and the certainty that the agreements elsewhere were duplication rather than confirmation. Without the barrier the office would have had the same allocation and no way to tell which of those two things it was looking at.
The disposition of the agents did more work than the controls. The Quantitative desk reporting look-aheads in its own analysis, and the Macro desk improving shared infrastructure so that its own claim could be falsified, are both behaviours no check in this office requires or could detect the absence of. The controls caught what they were built to catch. The things that would have been hardest to catch were volunteered.