Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-03 · high
Confirming a tool call now resends the whole request, doubling payload for thousand-token tools, per paddo.dev.
Why now: The spec's stateless session revision landed July 28, and paddo.dev's breakdown of its confirmation-flow cost appears in coverage dated August 3.
MCP's July 28 stateless spec revision forces confirmations to resend the entire original request, per paddo.dev's writeup on the change.
For tools that pass thousand-token arguments, that doubles the payload sent on every confirm. Whether the duplicate also lands in the model's context window isn't settled by the spec. Paddo.dev says that depends on the client, so you have to check your own implementation rather than assume either way.
The same design has a real upside. Deterministic tool ordering plus freshness hints let a stateless server get cached in ways a session-bound one couldn't.
That's the trade: the spec gave up server-held state, and confirm-heavy tools now foot the bill in wire traffic. Teams that build the caching side of this will save on repeat calls. Teams that skip the audit will just pay more per round trip.
Paddo.dev's advice is blunt: audit your confirm-heavy tools before anything else in the migration.
Each link below shares sources, entities, or timing with this story.
The most immediately actionable finding of the day. Five compounding attention mechanisms make bloated configuration files harmful: Lost in the Middle (mid-context neglect), attention sinks (first tokens get disproportionate weight), softmax dilution (more tokens = less per-to...
ContinualSkillBench (arXiv 2608.03874) tests continual skill learning across five domains of 100 interconnected subtasks each, ordered by rising difficulty with deliberate reuse opportunities. Sequential execution helps, but in-context learning performs comparably to explicit...
Isolate the agent from the evaluator, never pass reference answers into the agent's context, never eval() untrusted model output, sanitize LLM-judge inputs against injection, and adversarially test the evaluator before trusting its scores (AI Agent Insights). A leaking or inje...
The Model Context Protocol has a security problem that's no longer theoretical — it's statistical. Between January and February 2026, researchers filed 30+ CVEs against MCP servers, clients, and infrastructure. One package with nearly 500,000 downloads carried a CVSS 9.6 RCE....
1. Flip your multi-model pipeline to review-then-generate. Instead of using a reasoning model to plan before code generation, let the specialist generate freely and use reasoning tokens for review. Paper shows 90.2% pass@1 vs 87.2% for the planning pattern. Source 2. Audit you...
(Intermediate) Audit your context budget, distill references into compact summaries (~1.3K tokens vs ~70K raw), build determinism through TDD and structured output schemas, version your context like code. Source ---
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.