Cerevisor Insights

Research and perspectives on AI adoption for CEOs and funded founders.

New: AI Adoption Trend Dashboard. A monthly meta-view of every business-adoption post, with dual-verified figures and source links to every claim.

• Local reasoning models in a multi-agent workflow: where should the thinking go?

  Local reasoning models write out their thinking before they answer. In a chain of AI agents, that text can be passed along as if it were the answer, or use up the whole reply budget. Here is where it

• Cerevisor 3.9: AI agents that use websites and apps, and agent groups on a free canvas

  Cerevisor 3.9 lets AI agents use websites and desktop apps one approved step at a time, adds a free canvas where one message reaches a whole group of agents, and gives every agent a chat you can watch

• Reviewing AI agent output: why the first flaw you find hides the second

  Radiologists have known for decades that finding one problem on a scan makes the next one easier to miss. The same thing happens when a person or a second agent reviews an AI agent's work, and the fix

• Local AI models: should two of them share one laptop, or take turns?

  A large local model drafts, a small one reviews. I measured what it costs to keep both in memory on one laptop with no graphics card, what it costs to load them in turn, and which number turned out to

• Cerevisor 3.8: a proper chat for local AI agents, and help that sees your screen

  In Cerevisor 3.8.0, talking to an agent is a proper chat, and the built-in help, Need a hand?, now knows what is on your screen. How both work.

• AI agent evaluation: what an agent's stated confidence is worth

  Cerevisor asked every agent how sure it was, then let a high answer skip the second opinion. Two weeks of real runs and three studies later, here is what a stated confidence number does to the machine

• Cerevisor 3.7.1: an autonomous coding agent that works a queue

  Cerevisor 3.7.1 adds Factory mode: a workflow works through a list of things to build, one item at a time, has a separate agent check each result, tries again on a fail, and commits locally without ev

• AI agent governance patterns: how people approve, check and audit agent work

  A reference of thirteen patterns for taking AI agents to production: what a person sees at an approval, who checks the work, what gets written down when nobody is at the desk, and what an audit log ha

• How to Size vLLM Quantization for Multi-Agent Workloads

  Shrinking model weights with 4-bit quantization solves cold-load graphics memory, but concurrent AI agents immediately crash into context memory limits. Here is how to size vLLM quantization, cache bu

• Cerevisor 3.4 and 3.5: send a workflow as one file, then steer its agents mid-run

  Cerevisor 3.4 lets a workflow leave the computer it was built on, as a moved folder or as one shareable file. Cerevisor 3.5 lets a person talk to agents while they work. Here is what the two releases

• When the AI coding assistant hands the work back, who on the team can take it?

  Five Claude Code releases in three days were mostly about one thing: what happens when a person tries to interrupt an agent that is already moving. The comforting story is that a senior engineer can a

• Self-hosted AI agents: what bring-your-own-key actually protects

  A reference on running AI agents on your own API keys and your own machine: where the key actually sits, which of the four storage locations an attacker reaches first, and the three things local-first

• A Multi-GPU Box Won't Make Your Agents Run In Parallel By Itself

  New vLLM benchmarks show exactly what it costs in GPUs to serve real AI-agent traffic instead of a chat demo. Reading Cerevisor's own harness code shows why buying that hardware doesn't automatically

• Human-in-the-loop AI has a cost your dashboard doesn't show

  Cerevisor's new board feature rations how often one agent may interrupt another, but not how often an agent interrupts the person running the workflow. A 2008 study on interrupted work explains why th

• What updating to Cerevisor 3.0 unlocks, on top of 2.4

  Release 2.4 removed the four-agent ceiling, raising the scheduler's global concurrency budget to 64 concurrent cloud requests. Release 3.0, out today, builds directly on that foundation, making the ru

• The concurrency number that actually decides local-model GPU sizing

  Most GPU sizing conversations start with the wrong question. The real limit on how many agents can hit a local model at once has nothing to do with rate limits and everything to do with how the machin

• Failed, denied, or stopped: the AI observability gap in agent run logs

  Three of the four fixes in Claude Code 2.1.252 were about a failure being reported wrongly, not about work going wrong. That is the part of AI observability nobody prices: what a miscategorised red ro

• Where multi agent orchestration spends its time with the models switched off

  A reproducible measurement of what a multi agent orchestration layer costs before any model replies: nine cells, three run shapes, seven samples each, on one laptop, with the declared methodology, the

• What survives when a multi-agent system architecture loses the app mid-run

  Cerevisor 2.3.0 makes agent runs survive an app restart, and deliberately does not restart the agents. Here is the failure that forced the change, what the record now keeps, and why restored work stay

• AI agent orchestration frameworks report success. Almost none of them define it.

  Every tool that runs multiple AI agents ends a run with a status, and in most of them that status only means the process reached the end. Here is what a run result has to distinguish before any produc

• The attention budget your AI agent orchestration plan is missing

  Coding agents just stopped asking for permission, and the pitch is now fleets of them running while the laptop is closed. Research on attention residue says every delegated agent is an open loop in so

• How to set up an MCP server for a real agent workflow

  A reproducible guide to connecting an MCP server to an agent workflow: the six steps in order, working configs, what happens to the credential you paste in, the four failures I hit while building the

• Ollama tool calling is the local-model risk no benchmark measures

  Every open-weight model card reports reasoning and coding scores. None report whether the model can still ask for a tool correctly once it has been shrunk to fit the hardware you own. That gap is wher

• LLM pricing and capability matrix: the provider table I route agents against

  A maintained provider matrix for routing AI agents: context window, tool use, price per 1M tokens, latency, and MCP support, with every figure sourced to the provider's own docs and a published method

• Open weight vs open source: the distinction your board is about to get wrong

  A 1.6-trillion-parameter model shipped this week labeled open source, under an MIT license, with the weights initially pending. That gap is the whole lesson: open weight and open source are not the sa

• vLLM speculative decoding: the free speedup with a batch-size catch

  Speculative decoding can make a self-hosted model 2 to 3 times faster on the right traffic and 1.4 to 1.8 times slower on the wrong traffic. Here is how batch size and acceptance rate decide which one

• GGUF vs safetensors: which model format to download before you run locally

  GGUF or safetensors is not a quality contest, it is a question about what the file is for. Here is how to route by format, size the download, and test the quant you actually run.

• Which open-weight model to run when the release feed goes quiet

  No new open-weight model shipped this week, which makes the choice a stable one: license, VRAM fit, and your one task decide it, not the leaderboard.

• Ollama vs vLLM vs llama.cpp: which serving engine should run your model

  The serving-engine decision gets framed as a seniority ladder, and that framing is wrong. Ollama wins at low concurrency, vLLM pulls ahead under load, llama.cpp is the engine Ollama already runs, and

• How to run a small language model on a laptop without a GPU

  The open-weight release feed went quiet again, which turns local AI into a selection question instead of a shopping trip. Here is how to pick and run a small language model on the laptop already on th

• Your team adopted the coding agent. Did they adopt the delegated workflow?

  The first days of July, Claude Code and GitHub Copilot spent nearly every release fixing background agents that quietly died. That plumbing is the difference between a team that adopted a coding agent

• KV cache quantization: how far to compress before long context breaks

  Long context is expensive because of the KV cache, not the weights. Q8 and FP8 halve it for almost no quality cost, but Q4 quietly breaks the value cache and long-horizon tasks first. Here is how far

• The data-sovereignty question to settle before EU AI Act enforcement lands in one month

  Self-hosting an open-weight model changes who controls the stack and whose law reaches your data, not whether you are automatically private, license-clean, or cheaper. With one month to EU AI Act GPAI

• The week no model shipped is the week to fix your self-hosted LLM in production

  No open-weight model shipped in the last three days, and that is exactly why it is a good week to look at the operational stack around a self-hosted LLM. A fresh build guide and an Ollama release cand

• The self-hosted LLM to run is the one the VRAM tier already fits

  A fresh hardware-tier matching matrix lines up 2026 open-weight coders against the GPU a team actually owns. The model to self-host is set by the VRAM tier on hand and a real task eval, not the leader

• How to measure AI coding agent productivity after GitHub shipped a merge-count metric

  GitHub just added a merge-count-by-adoption-phase metric to the Copilot admin console. It is a great input and a dangerous board number. Here is how engineering leaders measure real coding-agent produ

• Self-hosted LLM cost vs the API: the utilization number your board never sees

  GPU rental prices barely moved this year, so the build-vs-buy decision is not really about the hourly rate. It is about one number most teams never measure: how busy the card actually stays.

• Cerevisor: A Local-First AI Agent Orchestration Platform for Multi-Agent Workflows

  Cerevisor is a local-first, provider-agnostic AI agent orchestration platform: a desktop app to visually build and natively run multi-agent workflows across major AI agent providers. Here is everythin

• GGUF vs AWQ vs GPTQ: which quantization format should you actually deploy?

  The quantization format for a self-hosted open-weight model is a hardware-and-task decision, not a leaderboard one. Here is how to read the GGUF, AWQ, GPTQ, and FP8 tradeoffs without getting fooled by

• How to Run an Open-Weight LLM in Production With vLLM

  Serving an open-weight model in-house is a sizing and operations job, not a download. A five-step playbook for budgeting VRAM, pinning a vLLM version, and running the break-even before you buy a singl

• Engineering hiring held up. Your org chart still changed underneath it

  New hiring data shows engineering is the most resilient tech job function in the agent era. That resilient headcount number is an average, and the average is hiding which roles grew and which thinned

• Self-hosting an open-weight model: the data-sovereignty question your board is actually asking

  Self-hosting an open-weight model changes who controls the stack and which jurisdiction reaches your data, not automatically your bill. Here is the calm version of the sovereignty conversation for the

• The Best Local LLM for Coding Is the One That Fits Your GPU, Not the Leaderboard

  The open-weight coding model topping SWE-bench is usually the one almost nobody can run on a single box. Here is how to pick a local coding model by fit to your task and your GPU, with the named model

• Self-hosting an open-weight LLM this quarter: who should own the memory?

  GPU rental prices are flat year over year while the memory you would buy to self-host is spiking hard. The build-vs-buy question this quarter is really about who absorbs the repriced asset, and a volu

• The LLM quantization myth: a good benchmark does not mean a safe quant for your task

  A quant that scores well on the leaderboard can still break the one task you ship. Here is why averages hide the failure, and the small eval that catches it before your users do.

• Self-Hosted LLM in Production: What Actually Breaks

  Self-hosting an open-weight model in production rarely breaks on the model. It breaks on the operations around it: KV-cache sizing, CUDA out-of-memory, cold-start latency, and the volume line below wh

• What this week's open weight models are quietly telling founders

  A read on the June open-weight release wave for technical founders: the frontier is moving weekly, but the headline models keep shipping without verifiable benchmarks and with memory floors no single

• GLM-5.2 Tops the Open-Weight Leaderboard. Your GPU Probably Can't Run It.

  Zhipu's GLM-5.2 open weights went live with a 1M-token context and a memory floor most teams can't run. Picking an open-weight model is a fit-to-your-GPU-and-task call, not a leaderboard download.

• What 'open weight' actually buys a regulated company, and what self-hosting really does for data sovereignty

  Open weights are not automatically license-clean, private, or sovereign. Here is what an open-weight LLM license actually permits a regulated company, and what self-hosting changes about whose law rea

• What LLM Quantization Actually Does to the One Task You Run

  Published quantization benchmarks make 4-bit look almost free, with average scores barely moving. The capability that breaks first is rarely the one the benchmark measures, which is why a 50-example t

• Your AI coding agent's automation just got its own meter. What breaks on June 16?

  On June 15, Anthropic moved automated Claude usage (the Agent SDK, headless runs, and GitHub Actions) off the flat subscription onto a separate metered credit, while GitHub shipped new controls for au

• Should you buy GPUs for local inference while memory prices spike?

  The memory shortage made GPUs more expensive to own, but a live tracker on June 15 shows cloud rental prices barely moved. For a board weighing local inference, the real question this quarter is not l

• How to Run an LLM Locally When the Best Model Won't Fit

  A new 1-trillion-parameter open-weight model landed this week that almost no team can actually run. Here is how to run an LLM locally as a sizing decision, so the model you pick fits the hardware and

• 5 questions that pick your AI coding agent (not 'which is best')

  The top coding models have converged, so 'which AI coding agent is best' no longer picks a winner. Here are five questions that choose the agent that fits how a team works, what it costs, and how chea

• When agents write the code, who owns verifying it?

  Tech layoffs hit their worst month in two years this week with AI as the top cited reason, and named companies are redrawing the engineering org chart around agent fleets. The box most reorgs forget t

• When one engineer can launch a thousand subagents, what does your engineering manager actually manage?

  Claude Opus 4.8's Dynamic Workflows let one engineer run up to 1,000 parallel subagents in a single session. The org-chart unit an engineering manager owns just shifted from headcount to human-plus-fl

• The AI coding adoption percentage on your CTO's slide isn't a productivity number

  A Jellyfish survey of 636 engineering professionals dropped a 64 percent productivity number this week. The same survey also dropped a different 64 percent. Both belong on the same board slide, and ri

• What is your June review actually grading now that the typing isn't the job?

  Q1 2026 layoffs hit 81,747 while Big Tech committed $725B to AI capex. Mid-year reviews start in about six weeks, and most engineering rubrics are still grading the work that harnesses now do. Here is

• The myth that AI is replacing engineers, in three numbers from this week

  Cognition just hit $25 billion selling Devin as the AI that replaces engineers. IBM rolled out its own coding agent to 80,000 engineers the same week. Here is what those two numbers, plus one from Mic

• Same Model, 26-Point Security Gap: The Harness Wrapper Decides

  A new benchmark showed one model producing a 26-point functional gap depending on the harness wrapping it. If your risk register names the model and skips the harness, it names the wrong unit.

Browse the archive (88 earlier posts)