---
title: "When Self-Hosted LLM Capacity Outgrew the Defaults Built Around It"
slug: weekly-recap-2026-09-12
date: 2026-09-12
excerpt: "Three Cerevisor releases and three posts this week share one honest thread: capacity grew somewhere, and the setting sitting next to it stayed the same size."
featured_image: "https://bbtxujdxvidaghmhxkqs.supabase.co/storage/v1/object/public/generated-images/blog-1789232652635-weekly-recap-2026-09-12.webp"
featured_image_alt: Abstract editorial illustration of a widening flow of parallel navy and gold data lanes running beside a single small steel gauge that stays the same size, representing capacity that grew faster than the safeguard next to it.
author: Kinan Hamwi
author_url: https://www.linkedin.com/in/kinanhamwi/
canonical_url: https://cerevisor.com/blog/weekly-recap-2026-09-12
updated_at: 2026-09-12T19:17:19.928941+00:00
---

# When Self-Hosted LLM Capacity Outgrew the Defaults Built Around It

### The week in one glance

- Four catch-up releases landed between September 4 and 11, from 3.0.1 up to 3.2.0, and the honest thread across all four is what stayed unchanged while everything else got bigger.

- The board feature rations how often one agent interrupts another. It does not ration how often an agent interrupts the person running it.

- Real vLLM benchmarks show what serious self-hosted LLM concurrency costs in GPUs. Cerevisor's own Sequential default was never updated to notice when a team actually buys that hardware.

## The theme: self-hosted llm capacity outgrew the defaults sitting next to it

This was a catch-up week, three releases and three posts that never made it into the last recap, and reading them together surfaces something none of them says outright: capacity kept growing, but only some of the settings next to it grew with it. Cerevisor’s concurrency ceiling went from four agents to a governed sixty-four, and the company shipped an audit trail to go with it. Good. But the same release cycle added agent-to-agent questions with a ninety-second timer and a three-question cap, then left questions addressed to a person completely unrationed. Separately, published self-hosted llm serving benchmarks show what real GPU concurrency actually costs, while Cerevisor’s own execution-mode default has not moved since the days when one modest GPU ran one model for one agent. Even the release pipeline itself hit an unrevisited limit this week, more on that below. The clearest evidence arrived last: 3.2.0 turned that same channel into a named feature, Question for you, an agent stopping to ask the person running it directly, with no timeout and no cap on how often it can happen.

## What shipped: four releases, and the pattern in what changed and what didn’t

  [**Cerevisor 3.2.0**](/releases/v3-2-0)
  release, 11 September

3.2.0 gave an agent a formal way to stop and ask the person running the workflow a direct question, called Question for you, with no timeout and ended only by an answer or by pressing Stop. The same release let a second, read-only agent review any agent's work before it counts, let a claim or decision carry up to eight pieces of evidence under Why believe this, and closes out a run the app died in the middle of instead of leaving it stuck claiming to be running.

  [**Cerevisor 3.0.1**](/releases/v3-0-1)
  release, 5 September

3.0.1 fixed the four bad days 3.0.0 had listed as still broken: a model call that never answers now times out with a plain reason instead of hanging forever, a shortened tool result is now marked as shortened, an unreachable connected server is now announced by name, and Operator spend interrupted by a crash now reaches that day's total instead of disappearing. It also gave provider reviewer blocks a one-click, remembered approval, and started labeling an agent that ran out of room to answer as ended early rather than finished.

  [**Cerevisor 3.1.0**](/releases/v3-1-0)
  release, 6 September

3.1.0 gave every run a board: a shared, live feed where agents leave notes for each other and for the person running the workflow, including a direct question to a named colleague with a bounded ninety second wait. A note is information, never an instruction, and it is written to the run that posted it, not to some wider store. That distinction is the release's own stated design rule, and it holds up.

  [**Cerevisor 3.1.1**](/releases/v3-1-1)
  release, 7 September

3.1.1 put the board's conversation on the canvas and let a new run pick up where the last one's board left off. The honest line is in its own verification notes: GitHub Actions has been unable to start jobs since September 6 over the account's billing, so this release was built by hand on Windows only, and the macOS and Linux update manifests were deliberately left untouched at 3.1.0 rather than offering a build that was never made.

## What we published

### AI adoption this week

  [**What updating to Cerevisor 3.0 unlocks, on top of 2.4**](/blog/cerevisor-v24-v30-what-you-can-do-now)
  scaling-operations

Release 2.4 removed the four-agent ceiling and raised the global budget to sixty-four concurrent cloud requests. 3.0 pairs that with a pre-run summary, live file tracking, and a run record that survives a restart, because wider parallelism without an audit trail was already a trust question before it got wider.

### AI coding agents this week

  [**Human-in-the-loop AI has a cost your dashboard doesn't show**](/blog/harness-human-in-the-loop-ai-interrupt-budget)
  harness-operator-sustainability

Cerevisor's board caps a question from one agent to another at three per run with a ninety second wait. A question addressed to the person running the workflow carries no cap and no clock. A 2008 CHI study on interrupted work found people compensate for interruptions by working faster, and the cost of that compensation, workload and stress, never shows up on a stopwatch.

### Running models locally this week

  [**A Multi-GPU Box Won't Make Your Agents Run In Parallel By Itself**](/blog/multi-gpu-box-agents-parallel-assumption)
  local-deployment

Fresh vLLM benchmarks show real self-hosted llm agent traffic, not a chat demo, needs deliberately tuned concurrency and tiered memory offloading to scale past roughly 64 to 128 simultaneous conversations. Cerevisor's own Sequential execution default and single generic stall timer were built for one GPU running one model, and nothing in the harness notices when a team has bought past that.

## Signals to implications for a self-hosted llm strategy

**Signal.** Cerevisor's global concurrency budget is a governed sixty-four concurrent cloud requests, and the accountability layer, pre-run summaries and durable run records, shipped in the same release cycle as the wider ceiling.

**Implication.** Before widening any automation's concurrency, confirm the audit trail exists first, not as a follow-up. *[Founder | Eng Leader]*

Source: [What updating to Cerevisor 3.0 unlocks, on top of 2.4](/blog/cerevisor-v24-v30-what-you-can-do-now)

**Signal.** Agent-to-agent questions are capped at three per run with a ninety second wait. Questions addressed to a human operator carry neither limit, and nothing counts how many pile up in a day.

**Implication.** Ask any harness vendor, including us, whether the product tracks a person's interrupt volume the same way it tracks a machine's. Most do not yet.

Source: [Human-in-the-loop AI has a cost your dashboard doesn't show](/blog/harness-human-in-the-loop-ai-interrupt-budget)

**Signal.** Real agent-serving concurrency needs configuration most self-hosted LLM buyers don't know to check, and Cerevisor's own execution mode defaults to Sequential regardless of what the GPU underneath can now do.

**Implication.** Before flipping any harness from Sequential to Parallel on new hardware, confirm the serving stack was actually tuned for the concurrency being requested. *[Eng Leader | Founder]*

Source: [A Multi-GPU Box Won't Make Your Agents Run In Parallel By Itself](/blog/multi-gpu-box-agents-parallel-assumption)

## The contrarian take on self-hosted llm defaults

A missing feature is visible. A default that never got revisited is not, and this week that is the more common failure. Cerevisor gave agent-to-agent questions a budget while leaving the channel aimed at a human operator completely open in [the board release](/blog/harness-human-in-the-loop-ai-interrupt-budget), and its own execution-mode default still assumes one GPU running one model even as [published self-hosted llm benchmarks](/blog/multi-gpu-box-agents-parallel-assumption) show what real concurrency actually requires. Neither is a bug. Both are choices nobody revisited once the thing next to them changed shape. The fix is not a bigger feature, it is a habit: every time a capacity number moves, ask out loud which limit was scoped for the old number and go check it by hand.

## Next week

If this recap was useful, the newsletter delivers it straight to your inbox every Monday. [Subscribe here.](/#newsletter)