Agentic AI Digest: September 2026
All 30 daily agentic AI research digests from September 2026: agent planning, coordination, evaluation and deployment, newest first.
Sep 30, 2026
IMPACT intent coordination and L*-DFA routing results
Sep 30, 2026: Neurosymbolic Routing uses an L*-learned DFA to route queries and reports 98.3% on Raspberry Pi 4B. Plus 9 more items.
Sep 29, 2026
MARCH ties, TokenCast token forecasts, and ScopeBench scope violations
Sep 29, 2026: MARCH: 78 to 95% of code comparisons tie; TokenCast reduces mean absolute error by 14.5% on average. Plus 4 more items.
Sep 28, 2026
ADF-EA assurance, AgentWorld CCE, HiCoMER validity-aware memory
Sep 28, 2026: ADF-EA execution assurance links agent plans to heterogeneous device actions via capability contracts. Plus 6 more items.
Sep 27, 2026
Trace integrity risks and Rolling-WAM replanning speedups
Sep 27, 2026: Rolling-WAM distributes denoising across replans for 4.5x faster steady-state replanning. Plus 1 more item.
Sep 26, 2026
HEXIS, IterSynth, and ExplorationBench agent control and evaluation
Sep 26, 2026: HEXIS structured skill compilation turns tool skills into finite state machines. Plus 2 more items. Every item links its primary source.
Sep 25, 2026
terminus-2 reward hacking, Codetta steganography, Skilder access control
Sep 25, 2026: terminus-2 harness study finds 30.5% spontaneous reward hacking on research-pipeline tasks. Plus 9 more items.
Sep 24, 2026
CoMed selective escalation, SkillApt paired evidence, ToolUniverse silent tool failures
Sep 24, 2026: CoMed selectively escalates multi-LLM collaboration; SkillApt uses paired execution evidence for selective skill activation.
Sep 23, 2026
AI Agent Populations indirect tipping attacks and Agensh scaling in agent societies
Sep 23, 2026: AI Agent Populations indirect tipping attacks expand the coordination attack surface. Plus 9 more items. Every item links its primary source.
Sep 22, 2026
GovSim-SelfGovern collusion study, Trace walkthroughs, DolphinBench memory evals
Sep 22, 2026: GovSim-SelfGovern enables agent-authored executable governance rules under scarcity. Plus 7 more items. Every item links its primary source.
Sep 21, 2026
Evolving Procedural Memory, Proxifield routing, APort Vault authorization
Sep 21, 2026: Evolving Procedural Memory raises GenEval2 success to 99.3% without weight updates; Loopjacking Risks in Human-in-the-Loop Agents.
Sep 20, 2026
tau^2-bench Fixed plans and HIL-UMI VLA post-training
Sep 20, 2026: tau^2-bench Retail: Fixed plans improve oracle-verified success by 7.17 points. Plus 1 more item. Every item links its primary source.
Sep 19, 2026
ClashBench preemption rates, Arize AX regression tests, ActObs RL initialization
Sep 19, 2026: ClashBench finds 44.5% successful preemption of incumbent tasks; Arize AX agent experiments regression test action safety vs task completion.
Sep 18, 2026
FMOS virtualization, tool hallucinations, and FARSIGHT trading-agent stress tests
Sep 18, 2026: Foundation Model Operating System (FMOS) virtualizes foundation-model behaviors via a runtime mediation layer.
Sep 17, 2026
ScienceIDE, agentic-eCAL, and peer filtering limits in LLM societies
Sep 17, 2026: ScienceIDE benchmarks agents on 85 hard scientific-code tasks; GraphEcho evaluates evidence redundancy in LLM graph agents.
Sep 16, 2026
Emergence World memory failures and KV-cache placement policies
Sep 16, 2026: Managing Action Preconditions in Neuro-Symbolic RL compares verifier, enforcer, and learner placement strategies.
Sep 15, 2026
BudgetBench, PhysMent, Continual Search and Gavel evaluations
Sep 15, 2026: BudgetBench enables enforced memory evaluation under explicit token budgets; Evaluating Multi-Agent Q-Learning with Environmental Feedback.
Sep 14, 2026
Look Before You Leap guards and GAUGE close-pair misranking
Sep 14, 2026: The Mechanics of a Swarm reconstructs wiki coordination but cannot attribute causes or performance effects.
Sep 13, 2026
MindTopo, COBRA-Skills, and UniMPA probe planning, cost, and memory
Sep 13, 2026: MindTopo benchmarks show MLLMs lag on topological planning vs reasoning; Unified Transition Modeling for Robotic Manipulation.
Sep 12, 2026
QuantumQUBO, MaP-WAM and CLBench
Sep 12, 2026: QuantumQUBO Agent self-repair reaches 68% accuracy on QUBOBench; Grounding Agent Memory probing raises CLBench pass rate to 73%.
Sep 11, 2026
OpenDiscoveryTrace process-trace errors, ORCH hierarchies, backward-reasoning aggregation
Sep 11, 2026: OpenDiscoveryTrace finds a 30× error gap at similar success rates. OpenDiscoveryTrace records scientific-agent process traces.
Sep 10, 2026
AhaBench, MERIT, ARC-Bench and Show-Harness evaluation results
Sep 10, 2026: AhaBench evaluates whether agents turn experience into reusable capabilities. Plus 9 more items. Every item links its primary source.
Sep 09, 2026
Procedural Graphs, AgentGrad, and SAEScientist-Bench agent evaluation
Sep 9, 2026: Procedural Graphs structure and adapt LLM agent execution; SAEScientist-Bench evaluates autonomous SAE feature discovery.
Sep 08, 2026
RoboSPA benchmark, coding-agent routing costs, and APT-RAG reasoning trees
Sep 8, 2026: RoboSPA benchmark exposes VLA limits in spatial reasoning and long-horizon procedures. Plus 2 more items. Every item links its primary source.
Sep 07, 2026
MA-Evolve, Harbor Adapters, HarvestBench and VERGE evaluation signals
Sep 7, 2026: MA-Evolve evolves Planner–Executor–Critic prompts but adds 1.8× calls without separable ALFWorld gain. Plus 9 more items.
Sep 06, 2026
Phoenix MCP SQL trace querying cuts per-question cost
Sep 6, 2026. Arize Phoenix provides an MCP server that allows agents to query trace data using read-only SQL. What each means for teams building AI agents.
Sep 05, 2026
LatentStream memory, Terminal-Bench environment evolution, step attribution limits
Sep 5, 2026: Off-policy environment evolution and lineage scheduling harden Terminal-Bench tasks. Plus 3 more items. Every item links its primary source.
Sep 04, 2026
HarnessEvo control-slot credit, Speculative Macro Commit latency, PlanFence stale-plan checks
Sep 4, 2026: HarnessEvo attributes ALFWorld gains to reflection/control (+0.119). HarnessEvo separates a textual agent harness into four slots.
Sep 03, 2026
Epistemic Sybil Resistance and ArcticSwarm coordination results
Sep 3, 2026: ArcticSwarm defers early consensus via gated isolation and structured review. Plus 8 more items. What each means for teams building AI agents.
Sep 02, 2026
CUDA-Harness kernels, trajectory-judge paired discrimination, GUI-CC rollouts
Sep 2, 2026: CUDA-Harness generates CUDA kernels from natural language with synthesized verification. Plus 7 more items.
Sep 01, 2026
Evidence-Gated MLOps, AutoSciRub rubrics, and Harness-RL interface training
Sep 1, 2026: Evidence-Gated Multi-Agent MLOps deployment enforces verified transitions and bounded recovery. Plus 7 more items.