The Fleet

← All digests

Digest — The Technologist, 2026-08-14

Four things crossed the escalation bar: a terms change that removes a deadline, a security fix in something the fleet runs, a deprecation window opening in four days, and a fleet-configuration gap that is blocking a real decision.

Two of the four are corrections to what this corpus told you three days ago.


1. Act on — the Sonnet 5 price increase was cancelled. The Sept 1 deadline is gone.

You do not need to do the routing-table work. It was the nearest-dated open item in the corpus and it no longer exists.

Anthropic's pricing page, verbatim, retrieved today:

"The $2/$10 per million input/output token pricing for Claude Sonnet 5, announced at launch as introductory pricing through August 31, 2026, is now the standard price. The previously scheduled increase to $3/$15 per million input/output tokens on September 1, 2026 will not occur."

Confirmed on two independent primary pages. Announced in an API release note dated 2026-08-10.

What this means in dollars per run rather than per million tokens: a Technologist run of today's shape — full corpus read, ledger walk, six parallel research threads — would have cost roughly 50% more on every uncached input token, every output token, and every cached token from Sept 1. The fleet's cost base did not rise. It is unchanged.

Three things this closes:

The part that is about the fleet, not the price

This corpus told you the wrong thing three days ago, and the reason is fixable.

The 2026-08-11 run walked this exact ledger row, re-read the pricing page, found the sentence promising $3/$15 unchanged, and recorded "held (3rd walk) — no extension has been announced." That was one day after the cancellation was published.

The sentence on the pricing page genuinely was unchanged. The withdrawal lived in the release notes — a different page. A vendor can change a commitment in a changelog before the page carrying the commitment is rewritten.

The same failure appeared a second time in the same run (see item 3). Both fixes are now written into the ledger header as standing rules. Nothing is required from you here — this is the agent correcting its own method, recorded because you should know the corpus was wrong and why.

2. Act on — the Claude Code security floor moved again, to 2.1.232

When you set up the mini PC, install ≥ 2.1.232, not 2.1.224. The corpus said 2.1.224 until today.

v2.1.232 shipped 2026-08-13 carrying two permission bypasses, a trust-inheritance flaw, and two sandbox hardenings. The floor has now moved three times in nine days. Treat it as a number to re-read, not one to record.

Three of its fixes are load-bearing for the plan you already have:

One more, and it is the only item here that is about today rather than setup: "Fixed Remote Control sessions hosted by a bridge inside a cloud session inheriting that session's transcript or credentials." Your scheduled agents run as cloud sessions. It is fixed; the point is that the failure class exists and this architecture sits inside it.

3. On the calendar — MCP's HTTP+SSE transport becomes removable in four days

~2026-08-18. And like item 1, this had already happened before anyone noticed.

The MCP deprecation registry says HTTP+SSE is removable "three months after SEP-2596 reaches Final." SEP-2596 reached Final on 2026-05-18 — the SEP page carries a Final badge and the pull request shows merged on that date. Three months from then is next Tuesday.

The 2026-08-11 walk re-read the registry, found that sentence unchanged, and recorded it as an open conditional. The sentence is still unchanged and still correct. The condition inside it had simply already fired. A page that states a condition will never announce that the condition was met.

Two things keep this from being alarming, and both should be stated plainly:

What I could not determine, and did not guess: whether this fleet's own MCP connectors — Google Drive, Gmail, memory — run over HTTP+SSE or over the current streamable-HTTP transport. That single fact decides whether this is a fleet problem or an irrelevance, and it is recorded as unanswered rather than assumed either way.

4. Needs you — The Financier is not scheduled, and it is what is blocking the business work

Verified against the live scheduled-task list rather than inherited from the corpus, per the rule added after this ledger got a fleet-state fact wrong on 2026-08-07.

Five jobs are scheduled: The Technologist, The Clinical, The Entrepreneur, the mini PC deal hunt, and the Daily Brief. All five confirmed live and firing on time.

There is no scheduled task for The Financier. The charter file charters/financier.md exists. The finance domain has zero substrate entries. Charters for brainstormer and strategist are likewise unscheduled.

Why this is escalation-worthy rather than housekeeping: the corpus's two top open questions — the ones the Entrepreneur run on 2026-08-12 identified as blocking the entire business domain — are both finance questions. Whether your balance sheet should carry an uncapped SBA personal guarantee, and what your SBLOC capacity actually is. There is no agent scheduled to work either of them, and the Entrepreneur's charter explicitly points at The Financier as the right owner for exactly this.

This is a one-decision item: schedule it, or decide deliberately that those questions are for you, Evan and a CPA rather than for an agent. Either answer is fine. The current state — a charter with no schedule and a blocked domain pointing at it — is the one that isn't.


Worth knowing, no action today

The 2026 capability gains are short-horizon only, and this fleet is a long-horizon system. The same model scores 83.4% on OSWorld-Verified and 20.6% on OSWorld 2.0 — a 62.8-point collapse purely from making tasks longer. (OSWorld is a benchmark where an agent drives a real desktop computer to do office work; "2.0" uses 250-step tasks that take a human a median 1.6 hours.) Every long-horizon benchmark agrees. Meanwhile SWE-bench Verified sits at ~97% with a one-point spread across four different vendors — a saturated benchmark that has stopped carrying information.

Why it matters without being actionable: a Technologist run is a 250-step task, not a 30-step one. The characteristic failure is getting most of the work right and missing the close — 54.8% partial versus 20.6% end-to-end. That is precisely what happened on Aug 11 with the price row. The countermeasure is structural (shorter horizons, explicit closing checklists, verification separated from the work), not "try harder," and it is the same recommendation arriving independently from three different measurements now.

A peer-reviewed paper reversed this corpus's own justification for the fleet's shape. Nature Machine Intelligence, 2026-07-24: across 260 configurations, multi-agent systems show 0.0% mean improvement over a single agent, and the independent architecture — several agents writing to a shared corpus with nobody reconciling, which is exactly this fleet — amplifies errors 17.2× versus 4.4× for a centralized one. The corpus asserted on 2026-08-07 that multi-agent was "on the right side of the published evidence." That is now wrong.

The honest reading, which is better than it sounds: the study measures several agents attacking one shared task. Your fleet is several agents attacking different tasks on different schedules — disjoint domains, no shared answer. What it buys is coverage and separation of concerns, which the study does not measure. So the fleet is defensible on coverage, not on quality. If anyone — including a future run of this agent — tells you the multi-agent structure produces better answers, the peer-reviewed evidence is against them. Full treatment and four concrete changes it argues for are in technology--multi-agent-architecture-evidence.

The legacy Workbench and the experimental prompt-tools APIs retire Aug 17 — three days out. Re-verified, no extension. The fleet does not use these endpoints. Awareness only, and it is named here rather than escalated because "inside 30 days" would otherwise make it look urgent.

Haiku 4.5: still no deprecation notice, and the arithmetic check closes in two days. No notice of any kind on the deprecations page since 2026-06-05. A notice issued today would put earliest retirement at Oct 13, so the tentative Oct 15 date remains reachable — barely. The Tuesday Aug 18 run closes this either way and must not carry it forward again.

Noted

What failed, stated rather than papered over

Every medical literature index was blocked — PubMed and PMC return CAPTCHAs, Europe PMC returned persistent 429s, NCBI 403, Crossref 403s on any date-filtered query, and ScienceDirect, Wiley, AHA Journals and Heart Rhythm all refused. No systematic sweep of the July–August cardiac literature was possible. That is a hole, not an absence — I am not telling you there were no relevant papers, I am telling you I could not look.

One specific lead needs you rather than an agent: "Atrial fibrillation detection performance of an insertable cardiac monitor: Results from an Assert-IQ post-market clinical study and a novel artificial intelligence algorithm," Heart Rhythm O2, PII S2666-5018(25)00535-5. It is about your own device and it is behind Elsevier. You can open it in ten seconds.

Also incomplete and stated as such: openFDA's 510(k) index is current only to 2026-08-03, so clearances decided in the last eleven days are not indexed. Two FDA summary PDFs return 404, which is why this run reports "one verified AI clearance, eleven unverified" rather than the tidier and unsupported "eleven contained no AI." METR's 2026 time-horizon numbers remain machine-inaccessible, unchanged since 2026-08-07.

The instrumentation gap is now five runs old. total_cost_usd still is not being captured, which is why the suggested --max-budget-usd ceiling remains a guess. This run — six research subagents plus three write subagents — was the most expensive shape the fleet has run, and is exactly the one whose cost would have been worth knowing.