Kakashi Venture Accelerator
The agentic corporate OS
Control8 puts a company's work, its agents and its knowledge on one operating layer. Work arrives on its own from the tools you already use, becomes memory you can question, and comes back as decisions.
Restricted to @ kakashi.ventures. No self-signup — new collaborators are pre-provisioned by an admin (Federico, Carlo, Alberto).
What it is
Not an admin tool with a chatbot bolted on: an operating system where the work of people and the work of agents run in the same space, over the same objects, against the same memory.
Where the time actually goes, what has stalled, what is worth doing now. Weekly planning, goals and blocks — plus the gap between what was planned and what happened.
Twenty-one scheduled automations and a registry of eighteen agents, each with its trigger, its decision and its off switch. More get added: the platform is where they work.
Everything the company writes becomes a corpus you can question: «8» answers with its sources attached, the graph tracks contradictions and open questions, and datasets and evals are built from the same material.
An MCP server inside the app, an Agent API for outside agents, webhooks and connectors. The structure doesn't stop at the interface: any compatible client becomes a surface of Control8.
How it works
Sixteen modules tell you what is there. This tells you how it connects — and it's why the memory doesn't go stale and the proposals aren't generic.
01
Coding sessions, merged PRs, email, meetings, documents. The work happens in the tools it was already happening in.
02
The plugin recognises sessions, Byakugan attributes PRs, intake routes the mail, the calendar becomes capacity. Nobody fills anything in.
03
Journals, wiki, dossiers and the knowledge graph rewrite themselves. The corpus answers plain questions, with sources attached to every claim.
04
Goals to confirm, suggested allocations, blockers queued, roster suggestions. Each one states why it is there.
The loop runs whether anyone is watching or not: capture is automatic, synthesis happens overnight, and the proposals wait for whoever opens their desk first.
What's inside
Each module answers a single question, and they are grouped by the question they ask. It's the same principle the menu is ordered by.
Your current work block, tasks, notifications and ready prompts. It's where you land every morning.
Questions the company's own corpus and answers citing what it rests on. If the data isn't there, it says so.
Person × day × block × project. Capacity comes from declared hours and from the real calendar.
by grant
Goals still alive and tasks still open: confirm or drop them, once a week.
by grant
The week at a glance: projects by family and by health, open blockers, goals in flight.
by grant
The shared unit of work: consulting, ventures, research streams and internal tools, each with its own header.
A queue of decisions waiting, ranked by severity. If there's nothing to decide, it's empty.
by grant
Where the work went, against where it was declared it would go. Analyses are saved and re-enter the corpus as evidence.
by grant
Wiki, journals and resources from every project on three axes that fill themselves in: nobody tags a link by hand.
What the company knows how to do: the skill library for Claude Code and Claude Desktop, with installs and usage.
Who's here, what they do, what they're on, and where to find them outside this.
The four modules the platform uses to measure and improve its own AI: Optimizing, Tuning, Training, Eval Setting.
by grant
The office from three sides: who buys from us, which rung the ventures are on, what we're learning.
Revenue, client roster, participation portfolio, campaigns and venture velocity.
by grant
Where we are, what we produced. Papers, findings and the trophy shelf.
Finance, legal, HR and office: labour cost, engagement margins, contracts, compliance across twelve frameworks.
by grant
How big it is
These are the platform's numbers today. They aren't hand-written: a test in the suite re-derives each from its source — the database schema, the migrations folder, the cron configuration — and turns red the moment one stops being true.
166
database tables
148
applied migrations
32
scheduled automations
71
MCP tools
138
pages
361
automated tests
A number in a document rots; an invariant doesn't. That's why you won't find a commit count or a green-build streak here: they're true as you write them and false a week later, and no test can notice.
The agentic side
An agent starts from a trigger that isn't a click, chooses between alternatives instead of applying a formula, and produces effects somebody must be able to review. They're the part of the platform that works when nobody is there.
Every project's «state of the art» dossier and the wiki rewrite at 2:15, the log→task linker at 2:30, the contradiction hunter at 2:45. Every run leaves a row in a ledger.
Contradictions don't self-confirm, proposed goals are confirmed, roster suggestions are applied by hand. What an agent writes unasked is only what has already happened.
The registry states, for each, what triggers it, what it decides, whether it observes or proposes or acts, which model it uses and which switch turns it off. Where no plan exists, the field stays empty.
Almost every writing agent has its own switch, and the two that don't say so. A switch that fails to switch off is discovered exactly when you need it: that's why there is one reader, and it's tested.
How it extends
Control8 exposes itself as a programmable surface. Any MCP client becomes a way of working in it, and any outside agent can read its state with its own credentials.
Twenty-eight tools inside the app: log activity, open tasks, set goals, fix a misfiled report, question the corpus.
The plugin recognises at session end whether it was company work, on which project, and writes one report per distinct deliverable.
The same integration for people on Codex: the server is client-agnostic, only how sessions are collected differs.
A read-only surface for outside agents, plus webhooks for portfolio, evals and communications. Dedicated credentials, explicit allowlist.
Nine chapters across every module, with links into the real pages. It opens after sign-in, because a tour of empty screens helps nobody.
Control8 · Kakashi Venture Accelerator · built in Torino · Notice and privacy
Who builds it
Control8 is built by the people who work inside it, and that's the first reason it works: whoever writes the code also uses those choices on Monday morning.
Adnan
Team Lead (ops)
Alberto
CTO - Co-founder
Carlo
Head of products - Partner
Davide
Researcher
Faezeh
Junior — UX
Giulia
HR - CEO PA
Giulia P
Neuroscientist
Manuel
Senior Writer
Maryam
Researcher
Matteo
Venture Analyst
Nicholas
Project Manager
This list comes from the platform's own roster, not a hand-written page: when the team changes, it changes by itself. Name, role and the external profiles each person chose to publish.
