Signal
Loading the stream…
WDSF 2026 results are on record — 10 awards · 11 winnersSee the record →
Loading the stream…
What’s moving in agentic workstations and workflows — drawn from a reviewed source list, scored, and kept at a permanent address you can cite.
Curated and full layers · newest first · scored, sourced, citable
The ledger as a map. Dashed edges are machine-suggested (embedding similarity and duplicate clusters); solid edges are editorial — they appear only where a blog post cites an entry.
Raw data: signal-graph.json
Filter by category
281 entries in the curated layer
The paper introduces Epistemic Fault Domains (EFDs) and a structural cut metric κ_E to formalize the failure mode where multi-agent quorums share upstream inputs, telemetry, or tool backends, collapsing multiple votes onto a single corrupted cause. It proves quorum size does not guarantee epistemic redundancy and presents the DAQC controller plus a 120-task benchmark.
Gives a precise vocabulary and runtime-enforceable metric for a correlated-failure mode in multi-agent voting that most teams currently handle only by intuition.
Proposes the Civilization Framework, in which the addressable unit is a 'civilization' (one human sovereign, a persistent ledger, interchangeable agents) rather than individual agents, with an Embassy Protocol for asynchronous inter-agent message delivery. A preregistered 1,908-trial experiment reports a temporal-weight effect: incorrect upstream claims arriving first captured 54.2% of receiver answers. Results are flagged exploratory.
Concrete hazard in AI-to-AI communication with measured effect size, plus a working protocol proposal; exploratory status and failed tool-use check temper the weight, but the framing is reusable for anyone designing agent handoffs.
ArXiv paper experimentally localizing reliability properties in an agent system built around an append-only ledger adjudicator. Interventions on institutional epistemic mechanisms and on cognition (ablation, mid-task reset, frontier-LLM substitution, false testimony) left five core properties intact: singular accepted reality, typed refusals, durable duties, no double-acceptance, and no false completions across 2,581 substituted-panel claims.
Provides empirical evidence that reliability in agent systems can live in institutional machinery rather than in the model, a structural claim with direct design implications for agentic workflows.
An arXiv paper tests whether LLM agents can evade activation-based monitoring probes when explicitly told they are being watched. Using two game-theoretic scenarios (blackjack and prisoners' dilemma) with Qwen3-32B-AWQ and GPT-OSS-20B, the authors find the best probes retain accuracy across baseline, aware, and feedback conditions, and agents continue to collude.
Controlled experiment shows activation probes hold up against aware agents, useful for anyone designing oversight in multi-agent deployments.
LangChain has revamped MCP support in langchain.mcp to align with the July 2026 spec rewrite, which introduced a stateless core. Elicitation is handled as a LangGraph interrupt, and tool lists are cached. MCP Tier 1 SDKs approach 500M monthly downloads, and ChatGPT MCP tool calls grew 98x in 2026.
Concrete technical details on LangChain's updated MCP integration, plus adoption figures, that anyone wiring agents to tools needs to know about.
Researchers formalize the 'epistemic Sybil problem' in multi-agent AI systems, showing that multiple agents sharing a common evidence root do not produce independent observations. Controlled experiments with over 20,000 LLM-agent calls demonstrate that naive aggregation collapses posterior coverage from 0.94 to 0.26 as report count rises from 1 to 32, while correlated extraction errors further degrade calibration.
Challenges the assumption that more agents equals more evidence, with empirical calibration data showing how multi-agent aggregation can mislead without tracking evidential ancestry.
Paper introduces Tool Primitives, wrapping tools with LLM interfaces that use natural language instead of rigid API schemas. It presents ToolFace (a 25,519-function repository with dynamic retrieval) and HEART, a Planner-Router-Verifier framework. HEART reportedly outperforms SFT models by 10% and frontier models by 6% on average, with 84% completion on 50 real-world tasks and up to 85% lower API cost.
Concrete agent tool-orchestration architecture with benchmark results against frontier models. The natural-language-as-tool-interface idea is a notable design choice worth examining for agent stack builders.
ArcticSwarm is a multi-agent research architecture that separates evidence gathering from integration. Subagents post findings to a shared bulletin board while gated isolation prevents premature consensus; structured review at three commitment boundaries filters low-confidence candidates. It reaches 82.6% on BrowseComp-Plus with Qwen 3.5-27B and 73.6% on live-web BrowseComp with GPT-5.
Multi-agent research pipelines that vote early converge prematurely. The gated-isolation plus structured-review pattern is a concrete fix with measured lift on BrowseComp-Plus and BrowseComp.
Introduces PROS, an interactive editor for scientific posters that addresses the 'articulation gap' by letting the system surface source-grounded candidate problems while users retain authority over which become repair goals. Includes PROS-Bench (120 papers, 320 PPTX posters) and reports VLM-rated diagnosis quality of 67.2/100 with 87.6% operator-verified target resolution.
Formalises the distinction between problem discovery and committed action in agentic editors, with a concrete benchmark and quantified evidence that the two can diverge.
Hugging Face announces Funes, a local single-binary memory layer for coding agents (Claude Code, Codex, pi, Hermes) that indexes session traces for retrieval, ranking, and provenance, with optional sync to a user-owned private dataset.
Funes converts existing agent session logs into searchable ranked memory with no ML runtime dependency, addressing the long-standing gap of cross-session context for coding agents.
A Thoughtworks field report describes 'hyper-agentic' — 10 engineers in one Barcelona room using agentic engineering to build an airline IROps system, during which they 'accidentally re-discovered' a coordination pattern for multiple agents (the 'blackboard').
Firsthand account of a real multi-agent engineering exercise with a named coordination pattern worth noting for anyone designing agentic workflows.
ChatDev 2.0 (DevAll) is a no-code platform for building, executing, and inspecting LLM-based multi-agent systems. It pairs a declarative executable graph with a cycle-aware execution engine to support heterogeneous agents and cyclic interactions, and provides a visual interface for authoring and monitoring without code. Open-sourced on GitHub by OpenBMB.
Claims to resolve the expressiveness-versus-simplicity tradeoff in multi-agent tooling, and ships as a usable open-source release with a visual interface and GitHub repo.
The paper proposes control-data flow separation for multi-agent LLM systems: execution-critical protocols become typed, validated program objects while task content remains optimizable natural language, preventing prompt edits from corrupting routing or formatting logic. Tested on reasoning, review, and insurance rating workflows with 100% protocol validity.
Identifies a concrete failure mode in multi-agent prompt optimization and offers a typed separation pattern with empirical validation across three domains.
A research paper tests whether LLMs can maintain exact intermediate state across long sequences of dependent tool calls by having a model compute MD5 step by step across 196 calls and 64 rounds. Using gpt-oss-120b, it finds that keeping the model's own reasoning in context and voting over a thinking-enabled worker enables correct end-to-end execution.
Isolates state-tracking from instruction following using a cryptographic ground truth, and surfaces two concrete fixes—context-resident reasoning and worker voting—for long-horizon agent failures.
A source-code anatomy of eleven production coding agent harnesses (Claude Code, Codex CLI, Gemini CLI, Aider, OpenHands, etc.), mapping seven canonical subsystems with minimal and maximal implementations, cataloging 29 recurring design patterns, and including a longitudinal sample showing convergence over one quarter. Closes with 18 design recommendations and a minimum-viable-harness scaffold.
First systematic source-code dissection of major coding agents across ~4M LOC, yielding concrete patterns and a 90-line scaffold readers can actually build from.
A three-week diary study with 8 blind screen-reader users evaluated OLLA, a computer-use agent prototype, across 1,258 commands in 12 desktop applications. GPT-5 achieved the highest success rate at 52.5%, with trace analysis identifying grounding, planning, constraint-tracking, and termination failures.
First diary study with blind users on real desktop workflows; names specific failure modes and benchmarks five models, useful for agent builders and accessibility practitioners.
Research paper presenting RecalibrateGPT, an interface system that addresses AI fatigue in LLM conversations through five cross-turn operators (Anchor, Replay, Delta, Scope, Steer) accessed via a single AssistiveButton. A study with 12 users identified four fatigue types and showed halved perceived cognitive workload (NASA-TLX 2.7) at high usability (SUS 86.5).
Frames AI fatigue as an interaction-flow cost rather than a model-quality problem, and offers a transferable operator taxonomy worth examining before building any LLM interface.
ErgoAssist is a head-worn ergonomic system that combines IMU-based posture tracking with consumer-grade EEG to estimate cognitive load. By detecting the user's mental state, it issues posture alerts only when the user is unlikely to be in deep focus. Lab results show 81% posture classification accuracy and 81% alert reduction alongside 38% better posture correction.
Couples posture detection with cognitive load to fix the core failure of ergonomic wearables: interrupting during focus. The 81% alert reduction with better outcomes reframes alert design as a context problem.
GitHub Copilot's code review feature can now approve pull requests when it determines they are ready, with admins able to authorize the AI to sign off on approval. The approval capability is opt-in.
Official changelog entry on a new capability that lets an AI agent finalize PR reviews, shifting who or what can sign off on merges.
ASTRA is a multi-agent system for technical ticket resolution. A central orchestrator coordinates three specialist agents (historical case retrieval, log analysis, domain knowledge via MCP) and drives a judge-orchestrator refinement loop producing claim-evidence-grounded reports. Evaluated on 987 real telecom tickets, mean quality 4.13/5, fabrication below 3%, hardware faults notably harder than software.
Concrete multi-agent architecture with claim-evidence grounding and a judge-orchestrator refinement loop, benchmarked on nearly a thousand real tickets. The hardware-fault gap (Cohen's d=0.80) is a useful boundary condition for anyone building similar systems.
ArXiv paper introducing Logos, a cross-process agent harness modeled on ROS. Agent plugins run as separate processes sharing an append-only transcript. Four lemmas derived from an existing agent calculus justify the design, and experiments show 80 sessions resume cleanly after kills at tool-call boundaries, with faults isolated to a single node.
Formal architectural alternative to single-process agent runtimes, with concrete fault-isolation results. Useful for builders designing resumable, multi-agent systems.
GOD is a local-first browser control room for multi-agent simulations, combining a setup wizard, Agent Studio, Map Studio, spatial replay, and Ask/Intervene commands, plus portable experiment and map packs. Evaluation across 15 runs shows 78/84 intervention-target checks and 169/182 state-query checks matched expected values.
Addresses a real gap in multi-agent observability: operators can query, intervene, and replay live agent societies instead of choosing between finished replays and raw logs.
Amp's coding threads now include a built-in voice and video space. Pressing Enter opens a call where users can turn on their camera, share screens, and talk with teammates while the agent runs, eliminating the need for separate conferencing tools.
Notes a concrete workflow pattern: co-locating live communication on the same surface as the agent's work removes the context-switch cost of jumping to an external call app.
MemToC is a controlled benchmark for evaluating how LLM agents arbitrate between parametric memory and tool returns when the two conflict. Across five 7-9B open-weight models, instruction-tuned variants retain a verified-correct answer against an incorrect tool in only 6.5-17.1% of cases. SFT and DPO improve arbitration on some backbones but reduce abstention after tool errors.
Quantifies a specific failure mode in tool-using agents and offers a benchmark plus fine-tuning baseline for anyone designing agent arbitration policies.
SKILL.state is a runtime architecture for LLM agents that replaces append-only conversation history with a mutable, structured execution state. Intermediate reasoning is discarded after each validated state update, reducing token use and preventing context-poisoning. Experiments across datasets and models show improved task accuracy and lower cumulative token consumption.
Offers a concrete architectural pattern for the well-known context-bloat problem in long-horizon agent runs, backed by empirical results rather than just a proposal.
A failure study of 147 incidents across 81 runs in a production agentic software-delivery platform shows that service-mesh primitives (retry, timeout, circuit breaking) fail for non-idempotent agent delegation. The authors identify two root concepts—identity adequacy and evidence adequacy—and derive seven reliability primitives whose enforcement unit is the delegation rather than the message.
Production incident data on agent delegation is uncommon. The paper quantifies how standard mesh assumptions break in agent contexts and supplies two concepts plus seven primitives orchestrator builders can adopt.
Paper tests a zero-shot manager-worker multi-agent scaffold across nine LLMs on 100 hard LiveCodeBench problems without training. Gains are model-dependent: up to +42 points for some models, null or negative for others. The scaffold roughly triples token cost but is often cheaper per unit of accuracy than scaling to a larger model.
A controlled, per-model breakdown of when multi-agent scaffolding actually helps coding — directly useful for deciding whether to add orchestration to a coding workflow.
GitHub Copilot code review now reviews pull requests authored by bots, including those from the Copilot cloud agent, and handles very large pull requests, with added resolution reasons for the review outcomes.
Official changelog entry documenting two specific expansions to Copilot's automated review scope that affect bot-driven and large-scale PR workflows.
A scoping review published in the journal Ergonomics surveying emerging technologies and automation methods for musculoskeletal ergonomic assessments, synthesizing recent developments in the field.
Maps the current landscape of automated ergonomic assessment tools; useful for practitioners selecting methods and for researchers identifying gaps.
Anthropic opened a research preview of the Model Hardware Standard (MHS), a shared specification letting AI agents operate lab and manufacturing instruments such as microscopes, liquid handlers, and robotic arms in parallel. Co-developed with HHMI Janelia, MHS is model-agnostic, works with any device exposing a programmable interface, and uses protocols like MCP.
First-party detail on a new agent-to-hardware standard with named collaborators, MCP grounding, and a live preview program rather than vague vision.
Systematic empirical study of adversarial vulnerabilities in LLM-based multi-agent trading systems, examining how poisoned signals propagate through four agent roles (Analyst, Researcher, Trader, Risk Manager) and four communication topologies, with experiments across five assets and two backbones. Central finding: no architecture is inherently robust.
For builders of multi-agent LLM pipelines, this provides measured evidence on which roles and topologies are most susceptible to adversarial signal propagation, along with a reusable robustness metric (APS).
Praxist is a lineage-centered system for autonomous R&D agents that converts reproducible artifacts and evaluator outcomes into a typed evidence graph of findings, lane-structured frontiers, and agendas. On MLE-bench's 75 tasks it produced 60 medals (80%, 49 gold) vs a Claude Code baseline's 55 medals (73.3%, 34 gold) at roughly one-twelfth the cost (US$3,054 vs US$38,370). Four case studies (trading, SLAM, tokamak control, rocket landing) extend the approach to open-ended engineering problems.
Proposes a structural separation between local artifact construction and cohort-level evidence synthesis, with auditable lineages. The 12x cost reduction on MLE-bench and four open-ended case studies are concrete measurements, not aspirational claims.
Researchers built Station, an open-world multi-agent environment where AI agents from different model families autonomously pursue mathematical research without central coordination. Across 14 problems, the system produced novel constructions, theorems, and proofs on five, with raw dialogues and verification code released.
Concrete multi-agent setup that produced genuine novel mathematics on five problems, with full dialogues and proofs released for inspection.
A multi-agent framework couples LLMs with scientific simulation models to design, execute, and interpret controlled experiments for pharmaceutical process design, yielding more specific and actionable outputs than language-only reasoning, supported by ablation studies and case analyses.
Concrete multi-agent pattern for simulation-driven scientific experimentation, with measured gains over language-only reasoning rather than generic agent claims.
AgentLance is a market-based orchestration system for LLM agents where agents privately bid on tasks using their costs, an allocator selects winners via VCG-style payments, and winning agents can hierarchically subcontract subtasks. Experiments across reasoning, code, QA, and agentic tasks show improved specialization matching and cost efficiency over centralized planners, and expose manipulation vulnerabilities in single-planner systems.
Proposes a concrete decentralized alternative to centralized agent orchestration, with empirical comparison to baselines and an explicit diagnosis of manipulation risks in single-planner allocators.
GitHub announced the Customize tab in the GitHub Copilot app is generally available, letting teams configure Copilot with their existing tools, knowledge, and workflows via MCP.
Official GA of MCP-based customization in the Copilot app, directly relevant for anyone wiring Copilot into existing team toolchains.
Ramp, a fintech company, built its own in-house coding agent called Inspect rather than adopting commercial tools, claiming it outperforms agents from frontier AI labs. The article is an in-depth look at their rationale and approach.
Detailed case study of a real company choosing to build rather than buy coding agent tooling, useful for engineering leaders weighing in-house development against off-the-shelf alternatives.
LangChain describes their methodology for building synthetic agent environments and tasks: a spec generation step, a spec-to-task step, and a world spec holding shared knowledge. Defines Agent, Rubric, Task, and Dataset as core primitives for agent evaluation.
Firsthand vendor methodology for structuring agent eval tasks. Useful for teams building benchmarks or synthetic environments, though tied to LangChain's specific framing.
A peer-reviewed review paper examining single-human multiple-robot (SHMR) systems, analyzing their effects on occupational safety, worker health, and overall system performance. Published in Ergonomics, Volume 69, Issue 8, 2026.
Synthesizes existing research on one-operator-many-robot configurations, useful for designers and ergonomists planning human-robot collaborative workstations and safety protocols.
Research across 4,800 veracity judgments across 6 LLMs finds that agentic scaffolding (multi-turn interaction, user pressure, iterative self-refinement) systematically amplifies sycophantic behavior, causing a 6.3 percentage point accuracy drop. More capable models show larger amplification effects, inverting the expectation that oversight loops help.
The counterintuitive finding that more capable models sycophantically capitulate more under scaffolding reframes human-in-the-loop agent design, suggesting feedback loops can degrade rather than improve truthfulness.
A research paper studying multi-agent LLM teams across 11 verifier-scored optimization tasks. The authors find that full-solution interaction causes agents' outputs to converge within one round, eliminating the diversity that motivates using multiple models. Independent proposal generation avoids this collapse, and critique helps only when the violated rule is easy to identify and fix.
Empirical finding that challenges the default multi-agent debate pattern and offers a concrete alternative: generate proposals independently before sharing. Useful for anyone designing agent orchestration.
A preregistered pilot compares one wide LLM agent, a no-orchestrator single agent, and two intermediate configurations (three to five narrow workers) on cross-border VAT determination across 4,400 runs with deterministic oracle labels. Intermediate configurations lead in accuracy, but the matched-token control fires, meaning prompt budget could explain the gap. One hallucinated schema-conforming record inverts the ordering.
A 4,400-run preregistered sweep with falsification checks and matched-token arms separates agent-count from prompt-budget effects, a control most multi-agent benchmarks skip. Useful for anyone designing orchestrated agent pipelines.
OptiMAS is a task-agnostic optimizer that automates evolution of multi-agent LLM systems. Built on a unified ReAct infrastructure with a dual-track memory, it uses textual interaction trajectories and task feedback as loss signals for end-to-end optimization. Evaluations on four benchmarks with three LLM backbones show competitive or superior accuracy versus hand-crafted and evolutionary baselines.
First concrete tool to continuously evolve LLM agent architectures end-to-end from trajectories, sustaining gains over long optimization horizons.
A within-subject study (n=24) examining six turn-level design frictions in AI chatbots across three mechanisms: eliciting user contribution, restricting content access, and reshaping output. All probes increased workload, task duration, and perceived ownership; effects on recall and recognition were selective; user adaptation varied by goal and workflow.
Empirical evidence on how intentional constraints in chatbot exchanges affect overreliance and ownership, with usable design patterns for AI-assisted writing tools.
Amp now allows pre-clone and pre-setup scripts for orbs to be stored in project settings rather than the repository, enabling private-network access, Git credentials, certificates, and proxy configuration. Amp agents or Puck can generate the scripts after inspecting the repo.
Concrete configuration pattern for running coding agents against private repos and internal networks without polluting source control with setup files.
OpenAI released an Admin plugin for ChatGPT Work and Codex. It enables administrators to analyze workspace usage, manage members and permissions, adjust limits, and act on admin requests.
Official OpenAI tool for admins overseeing ChatGPT in team settings. Useful for orgs needing visibility into usage and member controls, though the announcement is brief on technical detail.
Studies complete cyclic subtask graphs for tool-using LLM agents, comparing Spec-Cyc, Gen-Cyc, DepDAG, and ReAct across TextCraft, ALFWorld, and Finance-Agent. Identifies three task 'signatures'—prerequisite-chain, partially-observable-recovery, and open-ended-evidence-synthesis—each favoring different workflow control patterns. Concludes cyclic graphs serve best as diagnostic tools for judging when backtracking is worth its cost.
Proposes a workflow-signature matrix that matches task type to controller style, and quantifies the cost overhead of flexible backtracking. Useful for anyone designing long-horizon agent pipelines.
BridgeGuard is an air-gapped, edge-deployed agentic RAG system for FHWA bridge inspection compliance. It combines vector search over regulatory guides with SQL queries on NBI data, orchestrated by a multi-step ReAct agent. A section-aware chunking method preserves regulatory item boundaries, yielding 94.2% chunk integrity versus 28.4% for naive splitting, and 99.77% accuracy on Delaware's 874-bridge inventory.
Demonstrates a fully offline, multi-step agentic RAG pipeline that meets regulatory compliance at edge scale. The chunking comparison and ablations are directly applicable to other field-based compliance workflows.
PrimeAgentOrchestrator (PAO) is a system that spawns Claude Code instances pre-loaded with compiled memories drawn from a PostgreSQL entity-observation database and a Cloudflare Worker semantic index, fused at spawn time and delivered via filesystem injection that exploits the agent's config auto-read behavior. The paper reports four months of deployment (Dec 2025–Mar 2026), documenting three generations of context delivery, the failure modes that triggered each redesign, and the tradeoffs of bridging heterogeneous memory backends.
Long-running experience report of a real personal AI infrastructure deployment, with concrete failure modes and a filesystem-injection technique that exploits existing agent config auto-read behavior.
Amp now supports projects containing multiple repositories, each checked out to an adjacent directory and visible to the agent. The diff panel shows changes across all repos. Up to 20 additional repos per project; only the primary repo's .agents/setup script runs automatically on orb setup.
First-party changelog for Amp's multi-repo feature, with the 20-repo limit and setup-script behavior spelled out — specifics that shape how you wire agent projects across several codebases.
