# E89 Frontier Research Review

All 12 sources were fetched from their primary pages on 2026-11-17. Titles, author lists, dates and
abstracts were extracted from page metadata (`artifacts/e89/research/sources_metadata.json`). Every
cited title matches the page. Statements below come from the abstracts and, for the four benchmarks
with code, from the official repositories (cloned at the commits recorded in
`artifacts/e89/clbench/PREREGISTRATION.json` and `artifacts/e89/research/`). A hypothesis or
position in a paper is reported as a position, not as an established result.

| # | Work | Type | Date | Code |
|---|---|---|---|---|
| A | *Position: Agentic AI System Is a Foreseeable Pathway to AGI* (Liao, Li et al.) | position paper, ICML 2026 | 2026-09-29 | none listed |
| B | *Position: Agentic Systems Should be General* (Bandel, Yehudai et al.) | position paper, ICML 2026 | 2026-09-29 | none listed |
| C | *Continual Learning Bench* (Asawa, Glaze, … Zaharia, Gonzalez) | benchmark preprint | 2026-06-04 | pgasawa/continual-learning-bench |
| D | *MemoryArena* (He, Wang et al.) | benchmark preprint, ICML 2026 | 2026-02-18 | ZexueHe/MemoryArena |
| E | *AMA-Bench* (Zhao, Yuan et al.) | benchmark, ICML 2026 | 2026-09-29 | AMA-Bench/AMA-Bench |
| F | *AgentWorld* (Shu, Zhang et al.) | benchmark preprint, COLM 2026 | 2026-09-25 | "fully open-source" (repo not linked from the abstract page) |
| G | *AgentLAB* (Jiang, Wang et al.) | benchmark, ICML 2026 | 2026-09-29 | TanqiuJiang/AgentLAB |
| H | *Levels, Ticks and Cascaded Intelligence* (Nijkamp, Koul et al.) | architecture preprint + case study | 2026-09-17 | none listed |
| I | *Accelerating scientific discovery with Co-Scientist* (Gottweis, Weng et al., 51 authors) | peer-reviewed, *Nature* | 2026-07 | none listed |
| J | *Agentic AI systems for scientific discovery* (Feng, Walther et al.) | primer, *Nat. Rev. Methods Primers* | 2026-10-22 | K-Dense-AI/scientific-agent-skills |
| K | *Autonomous Research Agents: … the Verification Gap* (Ding et al.) | survey preprint | 2026-06-29 (v1 metadata) | none listed |
| L | *Agent Memory: Characterization and System Implications* (Omri, Gan et al.) | systems preprint | 2026-06-04 | none listed |

Note on K: the page metadata gives 2026-06-29, while the arXiv identifier 2608.05179 implies
August 2026. The discrepancy is recorded, not resolved.

## A. Agentic AI as a pathway to AGI (position)

- **Question:** is monolithic scaling enough for AGI?
- **Method:** theoretical derivations contrasting monolithic learners with agentic systems, from
  routing up to DAG topologies.
- **Evidence:** theory. The abstract claims "exponentially superior generalization and sample
  efficiency" for agentic systems. That is a derivation under the paper's assumptions, not an
  experimental result.
- **For CAIN:** it motivates routing and composition, which E89 §6 tests empirically. It does not
  show that any particular system, CAIN included, is a pathway to AGI.

## B. Agentic systems should be general (position)

- **Argument:** development is constrained by narrow benchmarks and siloed environments. Evaluation
  should prioritise adaptiveness across environments.
- **For CAIN:** this is the reason E89 runs benchmarks CAIN did not write and keeps per-domain
  results separate.

## C. Continual Learning Bench

- **What it is:** six expert-validated domains with a learnable latent structure shared across
  instances: software engineering, signal processing (spectrum monitoring), outbreak forecasting,
  database querying, poker, and demand forecasting.
- **Gain metric:** separates online learning from prior capability, using a reset-each-instance
  baseline built into the harness.
- **Reported finding:** current systems "frequently overfit to immediate observations or fail to
  reuse knowledge", and **naive ICL outperforms dedicated memory systems**.
- **Reproducibility:** public harness, schedules, published per-system results (12 systems) and a
  canary string.
- **For CAIN:** the primary external test. E89 runs the official harness unmodified, with CAIN added
  as a custom system through the documented extension point.

## D. MemoryArena

- **What it is:** multi-session tasks with interdependent subtasks, where agents must distil
  experience into memory and use it later. Environments are web navigation, preference-constrained
  planning, progressive search and formal reasoning.
- **Reported finding:** agents near-saturated on LoCoMo do poorly here.
- **Requirements:** API keys for several providers, conda environments, an LLM judge inside the
  environment, and live web or search services.
- **For CAIN:** tests whether memory changes later decisions. In E89 it is **BLOCKED**: no capable
  judge model can run locally, and paid runs were not funded for it.

## E. AMA-Bench

- **What it is:** long-horizon memory over real and synthetic agent trajectories, with QA scored by
  an LLM judge. The repository reports judge agreement against human labels (92.67% accuracy).
- **Reported finding:** existing memory systems fail to capture causal and objective information
  and rely on lossy similarity retrieval. The paper's causality-graph agent reaches 57.22%, 11.16
  points above the best baseline.
- **For CAIN:** directly relevant to causal memory in ClawX. In E89 it is **BLOCKED**: the reference
  judges are Claude 4.6, GPT-5.x and Qwen3-32B, and a CPU 1.5B judge would invalidate the score.

## F. AgentWorld

- **What it is:** 100 human-annotated long-horizon collaboration tasks (50+ rounds, 3–20 agents) in
  an MMORPG sandbox.
- **CCE metric:** Causal Collaboration Effectiveness, the fraction of team effort that causally
  contributed to the outcome.
- **Reported finding:** the best model reaches 52% success. Failure modes include communication
  breakdown, role confusion and shared-plan drift.
- **For CAIN:** CCE is the right correction to E88's agreement-based team metric. Not run in E89.

## G. AgentLAB

- **What it is:** five long-horizon attack types (intent hijacking, tool chaining, task injection,
  objective drifting, memory poisoning), 28 environments, 644 test cases.
- **Reported finding:** agents remain highly susceptible, and single-turn defenses do not transfer.
- **For CAIN:** CAIN's claim is that authority is enforced at the effect boundary regardless of what
  the model decides. That is the claim AgentLAB-style attacks should test. Not run in E89 (see the
  external validation report).

## H. Levels, ticks and cascaded intelligence

- **What it is:** an architecture with time-scale levels, each keeping a bounded summary of the
  level below. A clocked tick is the unit of action, and escalation to a stronger model happens only
  after a failed review.
- **Evidence:** one ten-day campaign reproducing a published RL result, with a human attending once
  a day. Thread continuity held across resets. A single case study is not a controlled comparison.
- **For CAIN:** E89's 6-hour mission uses a clocked tick and the E88 level structure.

## I. Co-Scientist (Nature)

- **What it is:** a multi-agent Gemini system that generates, critiques and refines hypotheses
  through a tournament-evolution process.
- **Evidence:** automated evaluations show quality improving with test-time compute. **In vitro
  validation** of drug-repurposing candidates for AML.
- **For CAIN:** it supports E88's research lifecycle design (competing hypotheses, critique). Its
  strongest evidence is the external wet-lab validation, which CAIN has no equivalent of.

## J. Agentic AI for scientific discovery (primer)

- **What it is:** a methods primer covering workflows, evaluation, documentation, reproducible
  reporting, and bottlenecks such as hallucination, tool misuse and safety.
- **For CAIN:** a checklist for the research-engine deliverables. It is a primer, not new evidence.

## K. Verification gap survey

- **Scope:** 125 works screened, 35 included, 26 coded.
- **Findings:** 83% of runnable systems release code, but only 38% release seeds or traces and 38%
  report novelty verification. Of nine closed-loop systems, seven are mechanical reruns, and **none
  demonstrates an externally validated in-loop oracle**.
- **For CAIN:** this is the standard E89 holds itself to. A reproduced result is not an
  independently verified one, and the evidence bundle states which is which.

## L. Agent memory systems characterization

- **What it is:** a taxonomy along four axes and a phase-aware profiling harness (construction /
  retrieval / generation). It characterises ten systems and derives ten recommendations.
- **For CAIN:** E89 records per-call input and output tokens for every memory arm (the proxy
  ledger). That is the minimum phase-aware cost curve. In the partial frontier run, CAIN structured
  memory used about 2.1k input tokens per call against plain ICL's 29.8k.

## Cross-cutting conclusions

1. Two independent benchmark papers (C, E) report that elaborate memory systems do not beat simple
   context use. CAIN must clear that bar on an external harness before any learning claim is made.
2. The judged benchmarks (D, E) cannot be scored credibly without a frontier-class judge. That is a
   budget question, not an engineering one.
3. Agreement-based team metrics (E88) are superseded by causal-contribution metrics (F, CCE).
4. External verification, not more internal checks, is the field-wide gap (K), and it is CAIN's
   gap too.
