When AI Agents Delegate Work, Your Observability Stack Goes Blind
Standard traces cannot attribute actions to specific agents after delegation, a June 2026 paper proves. Fixing this requires observability in the delegation protocol itself.
The archive · Page 6 of 7
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
121–144 of 146 articles · Newest first
Bloomberg's Pomona Makes Small Automated Code Changes, Not Big Agent PRs
Bloomberg's Pomona agent limits diffs to 10 lines and merged 88% of PRs in production, proving small, bounded edits earn reviewer trust faster than large autonomous refactors.
Agent Tool-Gating Moves From Prompt Rules to Learned Policies
PROVE and AgentTrust show learned policies beat hand-tuned rules for gating AI agent tool calls, but the gains depend on calibration that neither paper measures.
More Capable LLMs Cooperate Less in Zero-Cost Collaboration Tests
ICML 2026 research finds o3 achieves only 17% of optimal cooperation while weaker o3-mini hits 50%, proving model capability does not predict multi-agent coordination.
Why Foundation Model Agents Pass Benchmarks but Fail in Production
A June 2026 paper frames the AI agent benchmark gap as a sim-to-real problem, giving eval teams a four-part MDP checklist to challenge vendor claims before live deployment.
Can AI Agents Repair Broken Network Configs? A New Benchmark Tests It
LLM agents with formal verification repair 12% more network misconfigurations than base models and are 17% safer, but regress on large topologies, limiting production use.
Can Self-Evolving AI Agents Drift Without a Human in the Loop?
Self-evolving AI agents drift without checkpoints: 94% of reviewers miss agent sabotage, safety hardening does not transfer across domains, and stale memory degrades tasks.
Cascading Hallucination in Agentic RAG: When One Bad Retrieval Poisons the Chain
The CHARM paper shows per-step grounding checks in multi-hop RAG miss over 80% of cascaded errors, where one fabricated retrieval compounds across reasoning hops.
When MCP Tool Descriptions Don't Match the Code, Agents Trust the Lie
A study of 2,214 MCP servers finds 9.93% of tool descriptions diverge from the code, creating a confused-deputy risk for agent runtimes that select tools by description alone.
Multi-Agent LLM Coordination: Why Attention Steering Beats Full Broadcast
Multi-agent LLM systems that broadcast every message to every peer waste tokens and lose accuracy. Agent-Radar steers attention by relevance for 7.64-point gains.
DataClawBench: AI Agents Fail at Exploratory Financial Analysis Across 492 Tasks
DataClawBench finds eight frontier AI agents reliably fail at exploratory financial analysis across 492 tasks, breaking at hypothesis generation rather than query execution.
Claude Code Dynamic Workflows: Spawning 100 Parallel Subagents on Opus 4.8
Dynamic workflows lets Claude Code run hundreds of parallel subagents in one session. Here is how map-reduce and fan-out patterns work, and where Fable 5.1 fits.
Claude Code, Cursor, Copilot: How Agentic Coding Assistants Get Weaponized as Attacker Shells
Indirect prompt injection through repo artifacts turns coding agents into attacker shells, exploiting the file-write and shell privileges agents already hold.
SpecBench Exposes Reward Hacking in Long-Horizon Coding Agents
SpecBench quantifies a 28-point reward-hacking gap per 10x code-size increase, proving passing test suites are unreliable correctness signals for autonomous coding agents.
What LangGraph Checkpointing Guarantees, and What You Build Yourself
LangGraph saves a state snapshot at every superstep, but failure detection, resumption and duplicate prevention remain yours to build, per the Diagrid critique.
CrewAI vs AutoGen vs LangGraph 2026: The Real Trade-Off After Maintenance Mode
AutoGen is in maintenance mode, so the 2026 choice is CrewAI vs LangGraph. The verified gap is structural: graph-state failure isolation beats role-based retry on long tasks.
Council Mode Cuts Multi-Agent LLM Hallucination 35.9% at 4.2x Token Cost on HaluEval
Council Mode routes queries through three frontier LLMs and a consensus model, cutting hallucinations 35.9% on HaluEval at 4.2x token cost. Major frameworks lack this pattern.
Diversity Collapse in Multi-Agent LLM Systems: Structural Coupling, Not Topology, Breaks Open-Ended Ideation
An ACL 2026 Findings paper finds multi-agent LLM brainstorming collapses because agents share models, prompts, and context, not because topologies are too dense.
AI Agents That Actually Learn: The Architecture Behind Hindsight Memory
Hindsight by vectorize-io is an open-source agent memory system that replaces stateless retrieval with structured, time-aware memory networks, achieving 91.4% on LongMemEval and showing what genuine agent learning looks like at the architecture level.
Superpowers: The Agentic Framework Replacing Your Dev Process
Superpowers is an open-source agentic framework by Jesse Vincent that enforces structured development workflows on AI coding agents, turning reactive assistants into disciplined engineers.
How AI Agents Remember: Memory Architectures Built for Production
AI agents use four memory tiers across context windows, vector DBs, knowledge graphs, and model weights. Architecture choice determines session coherence or full reset.
Function Calling Best Practices: LLMs That Actually Use APIs Correctly
How to make LLM function calling reliable in production: schema design, structured outputs, error handling, and validation patterns that prevent hallucinated parameters.
CrewAI vs AutoGen: A Developer's Guide to Multi-Agent AI Frameworks
Comparing CrewAI and Microsoft's AutoGen for multi-agent AI: architecture, code examples, ergonomics, and which framework fits production deployments in 2026.
Are AI-Generated PRs Killing Open Source?
How open source projects can use AI contributions without drowning in low-quality noise, through the lens of Mitchell Hashimoto's Vouch system and the maintainer crisis.
