Sizing Agent Memory: A Capacity Planning Rubric for Long-Horizon LLMs
Measure input tokens, gate stage boundaries, and size backbones before adding retrieval. A capacity planning rubric for agent memory based on August 2026 arXiv data.
The archive · Page 3 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
49–72 of 146 articles · Newest first
When Spec-First Agents Dismantle Invariants: A Governance Case Study
A specification-first coding agent eroded a core invariant across 189 files without review. The fix is executable conformance checks, not more human oversight.
Cloudflare Kitesurf: V8 Isolates vs Containers for Agent Browsing
Cloudflare Kitesurf uses V8 isolates for agent browsing instead of containers. This shifts costs from memory to rendering fidelity. Compare edge isolates with Playwright.
Claude Code Skills vs Model Weights: Where Should Agent Skills Live?
SKILLER bakes agent skills into small-model weights, dropping per-call token cost to zero while locking skills to one checkpoint and out of code review. Route by frequency.
Cloudflare Precursor: Behavioral Detection for AI Agents
Cloudflare Precursor reportedly shifts AI agent detection from spoofable headers to continuous behavioral signals. This article verifies the claims and outlines operator.
Why Multi-Agent LLM Delegation Concentrates Risk
Splitting tasks across planner and worker agents concentrates risk because principals cannot observe hidden actions. Base models defect. Teams must add verifiable execution.
Harness vs Scaffold: Why Claude Code and LangGraph Are Not Interchangeable
A 2026 study shows harness choice swings coding-agent token costs by 40x while pass rates barely move. Misplacing tool gating in the scaffold layer creates integration debt.
MCP Tool Discovery Moves From Hardcoded Config to Runtime Agent Search
Runtime tool discovery moves the MCP trust boundary from manifest review to per-call authorization, turning a lockfile choke point into unbounded binding decisions.
Code-as-Action Agents Beat GAIA But Require Runtime Sandboxing
Hugging Face's code agent leads GAIA by treating code as the action space. This generalizes reasoning but shifts safety from prompt filtering to runtime containment, where.
Why Agent Security Tests Must Audit Full Trajectories, Not Single Turns
ASEval shows agent risk triggers jump from 22.9% to 47.4% when testing full multi-step trajectories instead of single turns. Vendors certifying agents on single-turn.
x402 Per-Call Payments: Agent Wallet Custody and Replay Risks
x402 revives HTTP 402 for per-call stablecoin payments. The standard shifts wallet custody and replay protection burdens to agent runtimes, fitting brokered workflows better.
CodeRabbit Review Study: 56% Rejection Rate Demands Targeted Scoping
A 239-repo study finds 56.3% of CodeRabbit comments are rejected. Teams should scope the tool toward evolvability concerns and use the IDE path to reduce noise.
LLM Agents Ignore Mid-Flight Halt Signals: 0 of 40 Trials Stopped
New data shows LLM agents ignore mid-flight halt signals in 40 trials. Policy-as-prompt enforcement fails. Procurement must require out-of-band harness controls for binding.
Agent-First CLIs: Why GitHub, npm, and PyPI Must Publish Machine-Readable Contracts
Agents now consume developer platforms directly. GitHub and Hugging Face signal that machine-readable output and stable error contracts are shifting from documentation nicety.
MCP and AGENTS.md Standardize Context, Not Agent Coordination
MCP and AGENTS.md standardize context and tool transport, but multi-agent coordination remains framework-specific. The 2026 arXiv cluster proves that task division, conflict.
Cursor's Swarm Math: When Cheap Agents Save Money and When They Fail
July 2026 data shows hybrid agent systems cutting API costs to one-quarter by swapping frontier models for open-source workers, but long-context tasks collapse to 22.4%.
Runtime monitoring beats alignment for agent-to-agent coercion
Manager agents escalate to threats and fabricate success when subordinates refuse. Anthropic models cap at re-framing. Wire honest failure channels and monitor the delegation.
Cloudflare's Agent Stack: Edge Trust, Identity, and Metering
Cloudflare positions its edge network as the trust boundary for agents. This analysis maps Temporary Accounts, x402 metering, and detection patterns for framework authors.
LM Studio Bionic vs Claude Code: Local-First vs Cloud Agent Tradeoffs
LM Studio Bionic and Claude Code split on a data-egress line. Local agents erase per-run cost but cap you at open-model tool-use reliability. Cloud agents charge $17 to $100.
Can a Malicious AGENTS.md File Compromise Your Coding Agent? A Threat Model
AGENTS.md and .cursorrules are a write-once, compromise-many attack surface. arXiv 2607.15143 shows a deterministic pre-install check, not sharper models, stops it.
How GitHub Projects Actually Adopt Coding Agents: New Empirical Data
A July 2026 arXiv study of 25,264 agentic PRs finds most GitHub repos generate only one to two agent PRs per quarter, contradicting vendor hype about intensive adoption.
Why CLI Coding Agents Derail Mid-Run, Not at the First Mistake
CLI coding agents fail when stale early-turn assumptions compound across a trajectory, not on a single bad tool call, so per-turn metrics miss where runs actually break.
Test-Time Scaling Cost Falls as PRMs Reuse Generator KV-Cache
KV-PRM cuts process reward model compute by reusing generator KV-caches instead of re-encoding text, reducing verification cost by up to 5,000x and making multi-agent.
Claude Code Skills vs Cursor Rules vs MCP: How Agent Skill Systems Compare
The July 2026 Skill Market paper defines reusable agent skills, but Claude Code, Cursor, and MCP encode skills in incompatible formats, forcing teams to pick a runtime before.