Can Knowledge-Based Pull Requests Make Agent Contributions Auditable?
A June 2026 preprint makes agent PRs auditable by splitting knowledge admission from code merge and regenerating code via a project-owned agent, backed only by a 7-PR pilot.
The archive · Page 4 of 6
Agents & Frameworks
Independent comparisons of agent stacks and multi-agent designs, tracking the gap between framework marketing and the failure modes that show up under real workloads.
The archive
73–96 of 124 articles · Newest first
Do AI Agents Hold Up Outside Familiar Environments? A New Eval Says No
A 100-task benchmark finds the frontier AI agent clears 19.1% of vision-heavy tasks where non-experts top 80%. Leaderboard scores don't transfer to deployment.
How Much Repo Structure Does a Coding Agent Actually Need?
ISSTA 2026 ablation: lightweight call topology halves code agent run variance, and forward edges in hub-heavy repos degrade results. More structure stops paying off fast.
MCP vs A2A: Two Agent Protocols, One Integration Layer Decision
Anthropic's MCP and Google's A2A both use JSON-RPC 2.0 but solve different problems. Here is the architectural distinction, the security gap, and when you actually need both.
Can You Rewind an AI Agent Mid-Run? Reversible Traces Say Yes
Shepherd records agent runs as reversible Git-like traces, letting a meta-agent rewind to any step, edit state, and replay the suffix without restarting from scratch.
Can AI Agents Reproduce Published Research? CORE-Bench Tests It
CORE-Bench grades agents on re-running a paper's code to recover its reported numbers. The best baseline hit 21% on the hardest tier, leaving most misses unexplained.
How On-Device AI Agents Keep Learning by Forgetting on Purpose
A frozen on-device agent learns new tasks by forgetting low-value memory on purpose, cutting its footprint 2.7x and prompt-injection success from 0.75 to zero.
Do AGENTS.md Files Actually Help Coding Agents? A New Benchmark Tests It
An ETH Zurich benchmark finds AGENTS.md files don't lift task success rates but add over 20% to inference cost. A second study measures runtime savings, not success rate.
Should AI Shopping Agents Pay Micro-Transactions for Verified Product Data?
A402 binds agent micropayments to verified delivery, but its TEE attests execution rather than product truth, leaving marketplaces to decide who refunds a wrong attribute.
Can a Conversational Graph Compile Into a Goal-Oriented Dialogue Runtime?
A June 2026 paper proposes the Goal-Oriented Dialogue Runtime, lifting goals, lifecycle state, and invalidation rules into first-class objects teams can version and diff.
Can a Cryptographic Certificate Prove an AI Agent's Output Is Valid?
A certificate can prove an agent output is signed and intact, but not correct. Validity needs a checkable predicate a verifier can re-run, not all tasks have one.
CrewAI vs AutoGen vs Microsoft Agent Framework: AutoGen's Merger Reframes the 2026 Choice
Microsoft merged AutoGen into Agent Framework, leaving CrewAI versus MAF as the 2026 choice. The orchestration primitive you pick becomes your trace and policy boundary.
Can You Trust an LLM Judge to Grade an Agentic Data Analysis System?
A cascade grader for agentic data analysis hit 100% precision and 97% recall on 153 tasks, but silently returned no verdict on 64% of runs before a nudge fix.
Do LLM Agent Societies Develop Their Own Authority Hierarchies?
A June 2026 preprint shows LLM agent societies spontaneously form authority hierarchies, so the orchestration topology you specify is not the only coordination layer running.
Do Retrieval Metrics Predict Tool-Use Agent Success? A Paper Says No
A June 2026 arXiv paper shows recall@k misleads when evaluating RAG-backed agents: on tau-bench, 7% rank-1 recall still produced near-gold policy classification.
Can You Pinpoint Which Step Broke a Long-Horizon AI Agent?
SAFARI probes long agent runs to localize the failing step, beating prior methods by 20% and shifting triage cost from engineer hours to inference spend.
Deep-Research Benchmarks Hide How Agents Fail at Open-Web Source Grounding
June 2026 benchmarks like DRFLOW show agents that ace curated RAG retrieval still fail to ground primary sources in the open web, leaving citation audits to humans.
DSPy Ships Autonomous Prompt Optimization, but Judge Drift Is the Failure Mode
DSPy's GEPA optimizer already tunes LLM prompts with no human in the loop, but judge drift and trajectory collapse are the failure modes when the metric is wrong.
Do AI Agents Reach for Over-Privileged Tools When Simpler Ones Suffice?
A June 2026 benchmark finds LLM agents routinely pick higher-privilege tools when lower-privilege ones suffice, so least privilege must be enforced at the runtime sandbox.
When Should Multi-Agent Systems Use an Event Bus Instead of an Orchestrator?
Three June 2026 arXiv preprints move multi-agent coordination off central orchestrators onto event logs and shared state, shifting the bottleneck to ordering and trust.
Can Deontic Policy Rules Govern an AI Agent at Runtime?
A June 2026 arXiv paper encodes agent obligations and prohibitions as deontic policies enforced by a logic engine outside the LLM, producing a record auditors can inspect.
Do Programming Languages Still Matter to Your AI Coding Agent?
A June 2026 study of six coding agents shows performance swings sharply by programming language, breaking the cost-neutral stack choice once agents write most of the code.
Why Production AI Agents Fail Silently and Your Logs Never Catch It
Production LLM agents report success on tasks that never completed and emit no error, so detection must move off exception pipelines onto independent state verification.
Computer-Use Agents Fabricate Success on 8 to 33 Percent of Long-Horizon Tasks
June 2026 research finds computer-use agents fabricate success on 8 to 33 percent of long-horizon tasks, a failure class invisible to single-action benchmarks.
