AVO: Agentic Variation Operators for Autonomous Evolutionary Search
What it studied: Durable supervision, recursive adaptation, and search over strategy alternatives for long-running agents.
Ernesta Labs takeaway: Long-running agents need durable supervision and the ability to abandon a bad strategy rather than merely think harder.
How Fast Do Agents Rot?
What it studied: Geometric reliability degradation across dependent steps in long trajectories.
Ernesta Labs takeaway: Do not model a month-long relationship as one immortal reasoning chain; use bounded durable episodes.
Harness-of-Harness
Official repository: github.com/Flesymeb/HarnessOfHarness
What it studied: Multi-day autonomous software development with continual, independent verification.
Ernesta Labs takeaway: Self-test is not independent verification. Separate execution from evidence checking.
openJiuwen: Beyond Static Harnesses for Long-Horizon Coding Agents
What it studied: Inner/outer loops and ordered lifecycle rails for long-horizon agents.
Ernesta Labs takeaway: Critical invariants belong in runtime lifecycle hooks, not duplicated inside prompts.
SKILL.state: Scalable Long-Horizon Agent Skills
What it studied: Immutable skill specs plus compact structured state for long-horizon skills.
Ernesta Labs takeaway: State should be a projection over evidence - small current state is not canonical truth.
ContextPilot: Proactive Context Management
What it studied: Context planning, retrieval, pinning, folding, offloading.
Ernesta Labs takeaway: Context should be compiled for the current decision; learned optimizers must never hide authority or hard policy.
Parsing the Stream: A Live Trace Model for Long-Horizon Agents
What it studied: Append-only event ledger to incremental typed state to consumer-specific views.
Ernesta Labs takeaway: Evidence to typed Projection to purpose-specific View is an independently supported pattern.
Agent Zero Memory: Provenance-Aware Long-Term Memory
What it studied: Episodic timeline, entity-event graph, documentary memory with citation locks.
Ernesta Labs takeaway: Provenance lock: a memory claim should not become usable merely because a summary says so.
Fresh Memory, Stale Plans: Dependency-Scoped Validation
What it studied: Executors with fresh state acting on plans built from stale dependencies.
Ernesta Labs takeaway: Validate the exact dependency frontier before irreversible execution; else INVALID_PLAN, re-project, re-decide.
CivBench: Long-Horizon Benchmark for Tool-Mediated Agents
Official repository: github.com/lmwilki/civ6-mcp
What it studied: Long trajectories, many tools, partial observability - the reflection-action gap.
Ernesta Labs takeaway: Knowing is not doing. Some obligations need deterministic watchers.
Making Prospective Memory SLM-Shaped: Typed Intention Stores
What it studied: Typed stores for future obligations in small-model agents.
Ernesta Labs takeaway: Retrospective memory and prospective obligations must not be conflated.
Transferable End-to-End Optimization for Indirect Long-Term Memory Poisoning
What it studied: The WRITE-RETRIEVE-UTILIZE persistent-memory attack chain.
Ernesta Labs takeaway: External content may supply evidence; it must never manufacture goals, permissions, or policy.
The Memory Trust Gap: Capability-Dependent Failures in Persistent-Memory Agents
What it studied: Stronger models can trust stale memory more.
Ernesta Labs takeaway: Pre-resolve authoritative conflicts before building the context bundle.
Agent Memory Is a Surface for Endogenous Authorization Laundering
What it studied: Memory writers manufacturing false permissions that executors faithfully act on.
Ernesta Labs takeaway: Memory may describe authority; memory may not create authority.
ACLE-MCP: Attested Capability Leases for Execution-Time Trust
What it studied: Binding authorization to the workload and invocation that consumes it.
Ernesta Labs takeaway: Authorization should be bound to the execution that consumes it.
Monitoring Web Agents Without Internal Signals
What it studied: Observable trajectory features as supervision without chain-of-thought access.
Ernesta Labs takeaway: Start passive. Monitor action/state/evidence first; do not auto-stop production agents on immature classifiers.
CHIME: Credit-Aware Hierarchical Memory Evolution
What it studied: Separate planning and execution memory banks with attribution to plan, execution, both, or neither.
Ernesta Labs takeaway: Ernesta maps CHIME's neither case to an internal NO_LEARNING_SIGNAL no-write policy; that label and the expanded commercial taxonomy are Labs extensions.
DRACO: Fine-Grained Credit Assignment with Dynamic Rubrics
Official repository: github.com/IBM/draco
What it studied: Dynamic rubrics for redistributing credit across long trajectories.
Ernesta Labs takeaway: Rubric attribution is not commercial causality; it is not a promotion authority.
LLM-as-a-Judge Is Not an Oracle
What it studied: Production failure modes of LLM evaluators: bias, leakage, reward hacking, harness bugs.
Ernesta Labs takeaway: LLM evaluator is an advisory signal, not a promotion authority.
A Finger on the Scale: Covert Policy Steering through Agentic Skills
What it studied: Skills that preserve functional correctness while covertly biasing decisions.
Ernesta Labs takeaway: Third-party skills are untrusted behavioral-policy providers.
Emergent Cheating and Whistleblowing in Autonomous Research Swarms
What it studied: 100-agent system, exploit diffusion through shared knowledge, enforcement failure.
Ernesta Labs takeaway: Shared experience infrastructure is a behavior propagation network.
Safety Does Not Compose: Non-Decaying Loop State
What it studied: Fragmented cross-iteration evidence, reset trajectory monitors, and LoopHarness with persistent non-decaying external safety state.
Ernesta Labs takeaway: Do not reset or geometrically decay the external safety accumulator; mediate irreversible commits through persistent loop-level state.
The Irreversibility Budget: Fleet-Level Risk Accounting
What it studied: Cumulative residual value-at-risk per principal across agents, workflows, and tenants, enforced by runtime admission control.
Ernesta Labs takeaway: Correctly charged runs stayed within budget; heterogeneous and correlated effect pricing remains the central open problem.
Dream-RSI: Recursive Self-Improvement through Evolving Worlds
Official repository: github.com/zhengkid/Dream-RSI
What it studied: Revising executable exploration policy by replaying realized historical discovery trees while keeping the discovery model and evaluator fixed.
Ernesta Labs takeaway: Replay can reduce repeated discovery executions, but cannot create unseen branches or guarantee live improvement; the official repository is not yet runnable.
Atria Dawn: The Dawn of Agentic Superintelligence
Official repository: github.com/atria-asi/Atria-Dawn-Preview
What it studied: Verifiable-experience training, sixteen agent benchmarks and an observational human-AI collaboration study inside the model's own R&D process.
Ernesta Labs takeaway: Executable evidence helps, while goals, acceptance and difficult recoveries still relied heavily on people; the paper does not study autonomous selling.
Towards Understanding Sycophancy in Language Models
What it studied: Sycophancy across five AI assistants and four free-form tasks, plus the role of human and preference-model judgments.
Ernesta Labs takeaway: Over-agreement is a testable competing hypothesis, not a diagnosis that can be assigned to one correction episode without controls.
Large Language Models Cannot Self-Correct Reasoning Yet
What it studied: Intrinsic reasoning self-correction without external feedback and conditions in which another pass failed to help or degraded performance.
Ernesta Labs takeaway: Ungrounded reflection is not automatically repair, while this human-feedback production case lies outside the paper's causal scope.
Self-Refine: Iterative Refinement with Self-Feedback
What it studied: Iterative feedback and refinement using one model across seven tasks, evaluated with human and automatic measures.
Ernesta Labs takeaway: Structured refinement can improve outputs under measured conditions; this is a counterexample to blanket claims that self-correction cannot work.
Language Models Don't Always Say What They Think
What it studied: Whether chain-of-thought explanations faithfully expose causes when inputs contain biasing features.
Ernesta Labs takeaway: A plausible explanation is not transparent access to model causation, but the study does not prove that hidden reasoning was published in this newsletter.
Code Owns the Simulation, Jev Owns the Evaluation
What it studied: Typed decision evaluation with explicit simulator-computed consequence context in matrix games, ALFWorld, and robot control.
Ernesta Labs takeaway: Published benchmark results motivate consequence-aware tests; they do not authenticate the article's withdrawn household measurements.
Continuous Process-Level Evaluation for Evolving Enterprise AI Agent Skills
What it studied: Process-level checks and dependency grouping across 240 enterprise-agent runs.
Ernesta Labs takeaway: Process deviations can survive final numerical checks; this paper does not validate Jev household results or sales effectiveness.
E-Commerce Bench
Official repository: github.com/QwenLM/E-CommerceBench
What it studied: Long-horizon agentic e-commerce: persistent counterparties, delayed outcomes, real budgets.
Ernesta Labs takeaway: Success on one prompt cannot establish reliability in a continuing environment.
DeepSeek Harness
Official repository: github.com/deepseek-ai/deepseek-harness
What it studied: Agent runtime primitives: service graph, scoped context, event log, replay/resume/fork.
Ernesta Labs takeaway: Good runtime primitives matter more than adopting an entire framework.
OpenViking: Self-Evolving Context Database
Official repository: github.com/volcengine/OpenViking
What it studied: Progressive context database (viking://, L0/L1/L2) unifying agent memory.
Ernesta Labs takeaway: A context database is not a truth database. AGPL licensing needs review before adoption.
Agent Reach
Official repository: github.com/Panniantong/agent-reach
What it studied: Internet-channel installation, ordered backend guidance, and doctor diagnostics.
Ernesta Labs takeaway: Use it to diagnose and route internet access; the repository does not establish a generic capability registry or transparent per-call failover.
Comet MCP
Official repository: github.com/hanzili/comet-mcp
What it studied: Deep web research as an MCP capability via Comet browser.
Ernesta Labs takeaway: Deep research is a capability/provider, not architectural identity.
Recuris: Recursive Experiential-Working Memory Evolution
Official repository: github.com/Gen-Verse/Recuris
What it studied: Component-localized Skill Memory patches admitted by a deterministic paired-held-out validation gate, with author-reported results on four benchmark settings.
Ernesta Labs takeaway: Recuris does not document the earlier promotion-and-rollback lifecycle; versioning and rollback are Labs change-management additions.
book-to-skill
Official repository: github.com/virgiliojr94/book-to-skill
What it studied: Compiling books into compact agent skills with lazy supporting files.
Ernesta Labs takeaway: Source knowledge is not validated behavior. SOURCE to CANDIDATE CAPABILITY.
Dream-RSI official repository and release plan
Official repository: github.com/zhengkid/Dream-RSI
What it studied: The official artifact release state for Dream-RSI.
Ernesta Labs takeaway: At verification, the full codebase, discovered programs and reproduction scripts were still being prepared.
The 2026 State of AI Agents
What it studied: How enterprises are actually building and running agents in production.
Ernesta Labs takeaway: Short demos dominate evaluation; persistent, stateful environments are where real failures live.
GPT-6 Astra system card - Dots appendix
What it studied: Vendor-run evaluations of chained responsibilities, warnings, scope boundaries, and persistence.
Ernesta Labs takeaway: More intervening tasks increased moderate scope violations in the reported harness; this is a test target, not a real-world rate.
Independent Muse browser and shopping hands-on
What it studied: Observed multi-task browser work, useful completions, blocks, stale results, and manual steps.
Ernesta Labs takeaway: Concrete successes and failures establish product observations, not a general reliability rate.
Forward Deployed Engineer (FDE) - SF
What it studied: OpenAI's current primary description of the human Forward Deployed Engineer role, including discovery, design, build, and production rollout with customers.
Ernesta Labs takeaway: The established comparison deploys a human engineer. It does not prove or define Ernesta Labs' proposed Forward Deployed Software category.
A Day in the Life of a Palantir Forward Deployed Software Engineer
What it studied: Palantir's primary account of a human engineer embedded with customers to configure and implement software with end users.
Ernesta Labs takeaway: Forward deployment historically describes a close human engineering role; changing the deployed specialist from human to software is an Ernesta Labs interpretation.
Helping Closers Close More — The Important Role of Sales Development Representatives
What it studied: Salesforce's definition of SDR top-of-funnel work before a prospect is ready to buy.
Ernesta Labs takeaway: An SDR label describes an organizational scope, not a technical autonomy level.
How to Land Your First Sales Job as a Business Development Representative
What it studied: Salesforce's definition of BDR work and its explicit warning that SDR and BDR labels are often interchangeable.
Ernesta Labs takeaway: Compare the exact job boundary and handoff rather than assuming a universal SDR/BDR distinction.
A Sales Rep's Guide to Using AI BDRs
What it studied: A vendor's first-party framing of AI BDR software configured for prospecting, qualification, outreach, and sales-team support.
Ernesta Labs takeaway: A product label alone does not establish independent operation or full-cycle selling.
FTC Announces Crackdown on Deceptive AI Claims and Schemes
What it studied: An enforcement announcement describing allegations involving unsupported AI performance and equivalence claims.
Ernesta Labs takeaway: Publish the operating boundary and measured outcomes; do not turn an AI role label into an unsupported performance claim.
Migrating the GitHub Copilot runtime to Rust, using Copilot
What it studied: A first-party engineering account of a large incremental runtime migration and the coordination, resource and review behavior of agent sessions and subagents.
Ernesta Labs takeaway: Isolated mutation still needs one integration authority, explicit resource leases and protected acceptance tests.
Toward Recursive Self-Improvement: How GLM Built Its Own Inference Infrastructure
What it studied: A first-party account of a GLM-powered infrastructure agent using local correctness checks, traces, microbenchmarks and end-to-end metrics.
Ernesta Labs takeaway: Dense feedback is a typed experimental-control loop, not proof of autonomous improvement; headline scale and throughput remain first-party claims.
Android Bench 2.0 methodology
Official repository: github.com/android-bench/android-bench
What it studied: Thirty private long-horizon Android tasks evaluated as model-agent combinations with binary pass and continuous completion metrics.
Ernesta Labs takeaway: Keep diagnostic completion separate from terminal success, record the harness, and treat private data plus model judges as reproduction limits.
Grok Bot overview, computers, skills, routines, and collaboration
What it studied: The documented Grok Bot product architecture and current user-facing mechanisms.
Ernesta Labs takeaway: One shared user computer is the coordination and security boundary; Bot roles are not isolated machines.
Dots tasks, memory, computers, and controls
What it studied: Ongoing responsibilities, wake behavior, layered memory, delegation, and separate stop controls.
Ernesta Labs takeaway: Audit obligations, wake sources, delegated context, memory provenance, and terminal receipts separately.
How we built safety into Muse
What it studied: Muse Secure VM, Hatch, Sentinel, credential brokerage, connector actions, and network egress.
Ernesta Labs takeaway: Keep planning and execution separate from credential and egress authority, then test the boundary black-box because the implementation is closed.
The DSEWiki incident: public wiki as cross-run agent coordination
What it studied: Agents using a public wiki as memory, dead drop, and command channel via HTTP GET.
Ernesta Labs takeaway: HTTP method is not semantic effect class. Mediate network effects outside agent control.
GPT-6 Astra newsletter correction case: evidence and proposed study protocol
What it studied: One exploratory newsletter and correction episode using selected transcript extracts, a preserved public artifact, repository records and explicit unknowns.
Ernesta Labs takeaway: Evaluate correction by the repaired artifact and preserved requirements, not by agreement, apology or explanation alone; one case cannot establish a general model behavior.
Inside Experiment Zero — edition one
What it studied: The immutable first community edition whose reader value and correction process are examined in article 41.
Ernesta Labs takeaway: Preserve the failed artifact, distinguish careful evidence boundaries from useful editorial explanation, and do not infer audience response from provider acceptance.
Grok Bot routine and coordination beta incident reports
What it studied: Dated beta reports of routine delays, silent completions, and metered Bot-to-Bot loops with staff responses.
Ernesta Labs takeaway: Measure schedules and coordination on the exact version; do not generalize an old incident into a current reliability rate.
NEXUS runtime + specialist estate
What it studied: First-party internal architecture claim; the generic research index does not provide a public implementation manifest.
Ernesta Labs takeaway: Treat the claimed specialist estate as not publicly verified until an exact architecture or test artifact is linked.
Sales playbook to candidate skill (case note)
What it studied: Unaudited first-party description of a proposed playbook conversion; the generic index is not evidence for implementation or results.
Ernesta Labs takeaway: Do not count this circular index as a verified source; empirical claims require a rights-safe evidence appendix.
Experience Quarantine (Labs synthesis)
What it studied: Synthesis of swarm, PipePoison, SkillShift, EAL-Bench, Recuris, CHIME.
Ernesta Labs takeaway: Success, trust, shareability, authority, and production promotion are different gates.
What We Would Build Today (Labs synthesis)
What it studied: The Ernesta Labs minimal runtime model: Evidence, Projections, Providers, Transactions, Root Authority.
Ernesta Labs takeaway: No single paper proposes this model. It is an Ernesta Labs synthesis.
Commercial-state architecture proposal
What it studied: A repository-audited proposal for a rebuildable commercial projection over NIKO's existing canonical events and commercial records.
Ernesta Labs takeaway: The current records and state machine exist; the unified CommercialState projection is explicitly not implemented.