Sixteen agents converged with no supervisor, and MCP overhead turned out to be scaffolding overhead

Sixteen headless Kiro CLI agents, one per EC2 instance, sharing nothing but an S3 bucket, converged on a written synthesis by iteration seven with no supervisor telling any of them what to work on. AWS published the pattern and the reference implementation yesterday as kiro-flock, Apache-2.0, and the interesting part is not the demo run but the accounting around it. Agents sit in a logical ring and each reads a fixed number of neighbors set by a radius parameter, so per-agent context stays constant whether the cluster holds 8 agents or 800; the largest system the authors have run totaled 184 agents across 11 cooperating clusters, building a programming language. Propagation in a ring is ceil(N / 2R) iterations and consensus runs two to three times that, which at a 30-second loop interval puts 1,000 agents at radius 4 around 62 minutes to propagate a single signal, or 12 minutes if you widen the radius to 20 and pay for the extra context.

The write-up is unusually honest about where the pattern loses. Full mesh visibility collapses the cluster onto whatever the first agent wrote, because every later agent reads that as consensus, so diversity dies at exactly the moment you wanted alternatives. Persistent session history produces drift, and the fix is counterintuitive enough to be worth stealing regardless of whether you ever run a flock: start every iteration with a fresh session and let the shared append-only logs be the only state an agent carries. AWS also cites the 2025 “Towards a Science of Scaling Agent Systems” result that architectures without centralized verification propagate more errors, and concedes the supervisor earns its bottleneck whenever you need a gate between steps rather than only at the end.

If you want a number that will change how you configure an agent this week, it came out of a controlled comparison instead of a blog post. The Scaffolding Matters More Than the Interface ran one fixed software task, six operations against a private git repo, across seven agent scaffoldings and five models, verifying completion by inspecting repository state rather than trusting the agent’s self-report. Two of the seven scaffoldings ship no MCP support at all, completed every run over plain CLI, and were 5.0x to 28x cheaper than the five that do support MCP, comparing CLI-only runs with no MCP server attached anywhere. A 27B model running locally varied 139x in cost across scaffoldings while finishing the task under all of them. The MCP-versus-CLI comparison the authors set out to make fell apart: thirteen strictly paired ratios span 0.43x to 29x. Where the two interfaces do separate is the cost of failure, 12.9% of money spent on MCP runs bought no completed work against 2.2% on CLI, though failures were equally frequent on both. Agents also routinely ignored the interface they were assigned, which means every practitioner cost estimate that did not verify actual behavior was measuring an unknown mixture. Harness, task, verification, and full dataset are open source.

A second paper landed on a file most of you edited this month. Why Does CLAUDE.md Keep Growing? traces 247,694 instruction lifetimes across 1,867 repositories and finds agentic prompt files grow +226% over their lifetime, gaining 4.9 net instructions per commit, with older instructions getting progressively harder to remove (log-hazard -0.032 per commit). The mechanism is asymmetric cost: appending is always cheap, but once an instruction’s rationale is gone, deleting it safely costs O(2^|D|) in a prompt of |D| instructions, so files stop growing only when someone rewrites them wholesale or the repo dies. The proposed fix is embarrassingly simple. Inverting IFEval to build worlds whose optimal prompts are known, comments that encode the latent reasoning behind an instruction removed 99.3% of the excess (+211.3% down to +1.4%), and the same trick improved real-world instruction-following on WildIFEval by up to 23.1%. The paper’s closing line is the argument: if English is the new code, why don’t we have comments yet?

Grab published Grab Bench, an internal eval harness built around a failure taxonomy rather than a leaderboard position. Their row-level analysis found SQL generation that keeps the query shape but changes the underlying metric, tool calls that pick the right tool family and drift on parameters, profile updates that cite every event instead of the supporting evidence, and coding agents that pass visible tests while missing a hidden stateful invariant. The design principle worth borrowing is the shortcut baseline: they run empty output, schema-only output, cite-all-evidence output, and no-op coding agents against their own scorer, and treat a passing shortcut as proof the benchmark is not ready. Deterministic scoring wherever the task contract allows it, LLM judges only for genuinely open-ended surfaces like SQL, hidden splits behind an access boundary so prompts cannot be tuned against the visible cases. They also state the limit plainly: synthetic evals show whether a model respects the contract under controlled pressure, not production uplift.

Tooling moved in smaller increments. Microsoft’s MAI-Code-1.1-Flash is rolling out across GitHub Copilot with native vision added to the small-tier coding model, and the previous MAI-Code-1-Flash retires across all Copilot surfaces on September 10; the same day GitHub shipped persistent Copilot memory and local Ollama models to the JetBrains plugin, and a per-model input/output/cache-read/cache-write token breakdown in the AI usage report, which finally makes credit consumption attributable. JetBrains argued the interop case directly in The “LSP Moment” for AI Agents, bringing the Agent Client Protocol to WebStorm so a team already paying Anthropic, OpenAI, or Google plugs that agent into the IDE instead of adopting whichever agent the vendor bundled. OpenAI made its Daybreak cybersecurity models available on Amazon Bedrock, a week after the expansion announcement, which puts them inside enterprise security workflows that were never going to call an OpenAI endpoint directly.

The loudest thing on the timeline was swyx calling a reasoning-trace extraction paper “already one of the most important papers of this year” and posting his own distillation because the methodology in the writeup is not clearly explained. Reasoning traces have been treated as a defensible moat for two years now, and the labs that hide them have been betting the hiding works.

What to watch: whether anyone reproduces the MCP overhead numbers with behavior verification attached, and whether a single vendor ships prompt-file comments as a first-class feature before the next CLAUDE.md rewrite cycle comes around.