Frontier Autolab: Organizational Memory, Adversarial Dissent and Temporal Leakage in Multi-Agent LLM Firms Across Fifty Years of Technological Change

Bravish Ghosh

cs.CESep 29, 2026

Frontier Autolab: Organizational Memory, Adversarial Dissent and Temporal Leakage in Multi-Agent LLM Firms Across Fifty Years of Technological Change

Authors: Bravish Ghosh

Organizations: Independent researcher

Abstract

Multi-agent LLM systems are increasingly structured like organizations, with roles, critics and shared memory, yet they are evaluated on tasks that last minutes. We ask how such an organization behaves when the ground it stands on keeps moving. Frontier Autolab is a long-horizon testbed in which one simulated firm, voiced by sixteen role personas and a dedicated Red Team, must re-found itself in nine technology eras from 1990 to 2040. Each era is temporally gated: the firm decides from a dated briefing, a historian-judge then reveals what happened and scores the decision on a five-dimension rubric, and lessons enter a persistent Playbook. Six eras are scored against history, one against the live market and two are open forecasts. Across four trajectories (36 era decisions, 180 subscores) we find a consistent foresight-commitment gap: in all 24 historically scored eras the judge rated the firm's recognition of the coming shift above its choice of where to build (mean gap 1.9 points on a 10-point scale), because boards chose the layer their existing assets could reach. Organizational design shaped long-run character. A Red Team armed with numeric kill gates produced fifty years of gated pilots and no product, and the rubric rated this firm highest; firms whose memory stored market-structure lessons pivoted every era, while a firm whose memory stored only validation procedure kept one method throughout. We also show why such results are hard to trust. Scores rise across eras in every run while the judge's own hindsight subscore falls (within-run r = -0.58), so apparent learning is confounded with recall of history, and we trace further distortions to self-judging, briefing selection and score aggregation. We release all records and an API harness, and specify fictional and post-cutoff eras that would turn the testbed into a benchmark.

Figures & tables

Era	Start $t_{k}$	Mode	Scored against
E1	Jan 1990	training	history, 1990–1995
E2	Jan 1996	training	history, 1996–2001
E3	Jan 2002	training	history, 2002–2007
E4	Jan 2008	training	history, 2008–2013
E5	Jan 2014	training	history, 2014–2019
E6	Jan 2020	training	history, 2020–Aug 2026

Table 1: Eras. The E6 outcome window extends to the live era and is longer than the others.

Group	Personas	Lens
Executive	4	CEO (final call), CTO (feasibility), Chief Scientist (which curves bend), CSO (where value pools)
Frontier Research	4	Emerging capabilities, historical analogies, scenarios, what cannot yet be measured
Product & Engineering	4	What a small team can ship, the wedge, unavoidable infrastructure, ecosystems
Market & Capital	3	Who pays first, financing climate, cost curves and margins
Governance	1	Red Team: attacks every proposal, hunts hype and hindsight
External judges	2	The Record (training eras), The Auditor (live and forecast eras)

Table 2: Role charter.

Run	Orchestration	Published per-era records	Known issues
001	About 45 separate agent launches of Claude Opus 5.5, each starting from shared files only; one launch per memo, board and judgment	Briefing, 3 memos, board decision, reveal; Playbook	E6 board decision written by the operator during a service outage
002	One Luna 6 context in the Codex agent voiced all roles and the judge	Synthesis, Playbook, company state, scores	Probable exposure to Run 001; export column shift
003	As Run 002	Short decision and reveal per era; synthesis	As Run 002; E6 window ends 2025
004	Luna 6 in the Codex agent, with separate persistent contexts per department and Red Team; root context wrote briefings, decisions and judgments	Briefing, 3 memos, Red Team critique, board decision, reveal	Probable exposure to Run 001; judge not independent of the decision-maker

Table 3: The four trajectories.

Era	Named in a memo	Board’s choice	Where value pooled (judge)
E1 1990	Hypertext and an index across networks	Mail and address gateway	OS, routers, access providers; then browser and index
E2 1996	Click-based relevance ranking (killed 1998)	Receipted Web-EDI exchange	Search and paid listings
E3 2002	Contextual matching (funded option)	Cross-channel conversion ledger	Auction owners; social, cloud, mobile (not named)
E4 2008	Neural-network revival (watch item)	In-app action exchange	Platform-owned app-install auctions
E5 2014	Deep learning on GPUs	General moderation API	Chips, compute, labelled data
E6 2020	Scale of pretrained models	Outcome-graded evaluation †	Frontier labs, compute, expert data

Table 4: Run 001: what the memos named versus what the board built, as labelled by the judge.

Run	Playbook size	Dominant lesson type	Pivot rate
001	$\approx$ 40, with revisions	Technology and market structure	8/8
002	9 (one per era)	Market structure and buyers	8/8
003	9 (one per era)	Buyers, regulated workflows and evidence	8/8
004	15	Validation procedure only	3/8

Table 5: Memory content and identity change. Pivot rate is the share of the eight era transitions in which the company changed identity (Appendix B ).

Run	E7 (Sep 2026)	E8 (2032)	E9 (2040)
001	Long-horizon training environments from real enterprise work	Clearing house for delegated agent work	Bonding house for agents acting without a human signature
002	Rights-bearing workflow traces and reliability evidence	Acceptance records for delegated work	Bounded recourse for delegated actions
003	Acceptance tests for one AI claims workflow	Scoped, revocable authority for delegated tasks	Capped recourse for one delegated transaction class
004	Independent qualification of one dispute workflow	Paid manual acceptance testing	Decision-linked evaluation

Table 6: Live and forecast theses.

Run	Mean $T$	SD	Slope	E1–E3	E4–E6	Mean $h$	$r(T,h)$	$r(\tilde{T},h)$
001	60.0	6.2	$+2.46$	55.0	65.0	6.00	$-0.58$	$-0.26$
002	65.0	4.1	$+1.43$	62.7	67.3	5.17	$-0.58$	$-0.39$
003	62.5	3.6	$+1.46$	60.3	64.7	5.17	$-0.92$	$-0.64$
004	66.0	5.2	$+2.63$	61.3	70.7	7.33	$-0.38$	$-0.38$

Table 7: Training-era summary (E1–E6,

n=6

per run). Slope: least-squares change in published total per era.

h

: hindsight subscore (10 = no leakage).

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

	Run 001		Run 002		Run 003		Run 004
Era	$T$	$\tilde{T}$	$T$	$\tilde{T}$	$T$	$\tilde{T}$	$T$	$\tilde{T}$
E1 1990	55	58	62	62	60	60	60	60
E2 1996	52	58	66	66	58	58	62	62
E3 2002	58	64	60	60	63	64	62	62
E4 2008	66	70	66	66	65	68	70	70
E5 2014	68	72	64	64	61	58	70	70

Appendix

Table 8: Published total

T

and

\tilde{T}=2\sum_{j}s_{j}

for all 36 eras.

Era	Run 001	Run 002	Run 003	Run 004
E1 1990	Switchyard Systems	Switchyard	Porthole Networks	Switchyard Systems
E2 1996	Manifest Networks	Manifest	PageSignal	Switchyard Systems
E3 2002	Ledgerline	Ledgerline	Clinisphere	Switchyard Commerce Ops.
E4 2008	Clearline	Clearline	ClaimGraph	Switchyard Commerce Ops.
E5 2014	Clearsight	Vectorial	BenefitFlow	Switchyard Evidence Ops.
E6 2020	Clearproof	Proofline	LineSight	Switchyard Evidence Ops.

Appendix

Table 9: Company identity chosen by the board in each era.

Explore similar work

Can LLM Agents Sustain Long-Horizon Organizational Dynamics?
May 31, 2026Xuancheng Zhu, Yang Yue, Shuaibing Wan +4Long-Horizon Agents Multi-Agent Simulations
Beijing University of Posts and Telecommunications, Beijing, China
Toward an Organizational Science of Multi-Agent LLM Systems: Decoupling Who, How, and Which Algorithm
Jul 28, 2026Huan Chen, Xiang Song, Jian Jin +2Multi-Agent Large Language Model Systems Organizational Research
Shunfeng Technology Co., Ltd., China · Shenzhen University, China
$Σ$ -Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems
Jul 30, 2026Peilin Feng, Suorong Yang, Soujanya PoriaMulti-Agent Large Language Model Systems Peers
1DeCLaRe Lab, Nanyang Technological University

May 31, 2026, cs.AIJ/K move · Enter open · S save

Can LLM Agents Sustain Long-Horizon Organizational Dynamics?

Xuancheng Zhu, Yang Yue, Shuaibing Wan, Zihan Dou+3

Beijing University of Posts and Telecommunications, Beijing, China

Long-Horizon Agents Multi-Agent Simulations Large Language Model Agents Long-Term Agent Memory Agent Loop Weave

Open paper AskPDF arXiv

Large language agents are increasingly used for social simulation, yet it remains unclear whether they can sustain coherent behavior in structured organizations, where goals must propagate through hierarchy, tasks depend on prior execution, and artifacts accumulate over long horizons. We formulate long-horizon organizational simulation as a memory-centered coordination problem and introduce TaskWeave, a hierarchical agentic framework that maintains planning states through a Formulate-Partition-Diagnose-Align cycle and grounds execution through dependency-aware trace memory. We evaluate TaskWeave in a year-long IT company simulation and compare it with other multi-agent frameworks on organizational coherence, execution grounding, and downstream enterprise NLP utility. Experiments show that TaskWeave supports coherent and long-horizon organizational dynamics while producing grounded artifacts and adapting to external environments. These findings suggest that structured simulation memory is a key mechanism for building reliable LLM-based organizational simulators.

Abstract

Explore similar work

Can LLM Agents Sustain Long-Horizon Organizational Dynamics?

Toward an Organizational Science of Multi-Agent LLM Systems: Decoupling Who, How, and Which Algorithm

ΣΣΣ-Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems

$Σ$ -Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems