At EDAMAME, my company builds systems that verify what software actually does and get better over time — runtime verification for coding agents, self-improving attack-pattern detection on the endpoint.

This did not start a few months ago. Four years ago we had a vision, and we made two early bets: we chose a single language, Rust, for everything we would build, and — we were lucky here — we met Michael Truell and the other founders of Cursor early enough to be picked as one of their first twenty users. We could see what was coming, so we built at the edge of it, our architecture growing alongside the AI tooling itself. At the end of 2025 we made the natural next move: we extended the same way of working to our go-to-market.

That extension is a system we call SIFU — our Self-Improving Founder Universal assistant. It is the self-improving go-to-market brain of EDAMAME, and I want to be open about how it works, because I think it is where a lot of companies are heading.

A go-to-market organism, not a chatbot

SIFU is not a single prompt, and it is not a seat in someone else's SaaS. It is a body of around 48 skill packages that live in our own repository and run from inside the editor. Each skill is a small, documented, testable capability — and together they cover most of what a go-to-market team does:

cold and warm outreach over email and LinkedIn, in the founder's voice, in English and French;
mining our own warm network for the right introduction path;
press pitches, blog posts, and the pitch decks;
investor updates, board emails, and the weekly momentum read;
product analytics, channel attribution, and even burn-rate analysis.

It is, in other words, a vertical specialist — the founder-and-go-to-market hemisphere of a company's brain — rather than a thin, horizontal assistant.

This is not new ground for us. The same Cursor-driven discipline runs across the workspaces that actually ship our product — the desktop and CLI app, the Hub, and the Portal — and it is why our speed to robust product updates and new features is, today, state of the art.

What that looks like, day to day

A few concrete examples make this less abstract:

Press. When a relevant article appears, SIFU drafts a pitch in my voice — English or French — pulls the right reporter from our own network, checks that every link in the email actually resolves, and leaves it in my Outlook drafts for me to read and send.
Warm introductions. Instead of spraying a cold list, it searches our own relationship graph — years of email across three prior companies, fused with LinkedIn first-degree connections and the live mailbox — to find the warmest real path into an account.
Analytics. Each week it reads our product analytics, web traffic, ad spend and CRM pipeline, then writes a single report that ties money spent to sign-ups actually earned — per channel, per segment.
Decks. It rebuilds the investor and customer decks from code, renders every slide to an image, checks them for overlaps and broken layouts, and only then promotes the clean version.
Finance. It turns a raw bank export into a burn-rate breakdown — R&D versus operations versus sales, including what we spend on the coding agents themselves.
Investor & board. It drafts the monthly board update and investor notes from the mailbox, the calendar and the pipeline — momentum, not vibes.

That breadth is the point: one shared body of skills covering the whole funnel — not a different, disconnected tool bolted on for each stage.

The self-improvement loop

What makes SIFU more than a pile of scripts is the loop it follows for every task. Before doing anything, it searches for a skill that already exists. If one does, it enhances it; if none does, it creates a new one. It executes the task. Then — the part that matters — it persists the improvement: it writes the new capability, a rule that teaches the system when to use it, and a test, and commits all of it to version control.

So the company's go-to-market knowledge stops living in people's heads and starts living in a versioned, testable, shareable body of work. SIFU is built on the open Agent Skills standard — plain SKILL.md files with progressive disclosure — that Anthropic, Cursor and others now share. The "soul" of the company becomes something you can read, review, and improve, not folklore.

One shared soul, many voices — kept consistent

SIFU is not one person's voice. The same shared brain is run by the whole team through Git: each seller works in their own identity and their own voice, in English or French, while drawing on one body of positioning, proof points and relationships. Many sellers, one soul.

That shared soul is also what keeps us consistent. There is a single source of truth for how we describe the company and the product, and SIFU keeps it in sync in both directions — across the soul itself, so positioning and knowledge stay current as we learn, and across everything the outside world sees: the website, the decks, the outreach, the press. When the message changes, it changes everywhere, on purpose — not three weeks later in one channel and never in another.

The opposite of a generic AI SDR

The market is crowded with company-agent ideas, and it is worth being precise about where SIFU sits. At one end are the AI SDRs — 11x, Artisan and the like — that promise to replace a sales team; most are shallow, doing one stage of the funnel from a bought database of a few hundred million cold contacts, and their output is routinely flagged as robotic. At the other end are horizontal company brains like Dust and Glean, and enterprise agent platforms like Salesforce's Agentforce and Sierra — powerful, but generic shells you still have to fill in. SIFU was built the other way around, on purpose: narrow and deep rather than broad and shallow.

Deep, not shallow. It runs the whole funnel, from first touch to investor relations, not a single outbound step.
Voiced, not generic. It writes in a real, specific voice — each seller's own, in English or French — with explicit rules for what that voice may and may not say.
Warm, not cold. It draws on years of genuine relationships to find the right path in, instead of spraying a bought list.
Measured. It is wired to our own analytics and pipeline, so it can tell whether a change actually worked — most agents have no idea.

And it stays honest. A human is always the first and the last step: a person sets the intent at the start, and a person reviews and sends at the end. Mass communications are created as drafts by default — the agent is deep and fast in between, but it never has the first word or the final one. That is not a limitation we are embarrassed by — it is the design.

From assistant to organism

I do not want to oversell where we are. A research frontier is racing to remove the human entirely — Sakana's Darwin Gödel Machine rewrites its own code and keeps the change when it scores higher on a benchmark; Voyager grew an executable skill library on its own inside Minecraft; Dust-style shared runtimes let intelligence compound across a whole organization in real time. SIFU is not that yet, and the distance is a roadmap I am happy to publish:

Close the loop. Because SIFU already owns its own outcome data, it can score each change against real results — reply rates, sign-ups, pipeline movement — instead of a synthetic benchmark.
Add a heartbeat. Turn the morning briefing and the watchers into a real orchestration layer, with budgets and approval gates.
Make the shared memory live, so context compounds across the team in real time, not just through Git.
Govern it for more than a tiny team.

Throughout, a human is the first mover and the last word. The goal is leverage, not abdication.

Where this goes: a sentient enterprise you can talk to

Step back far enough and a bigger shape appears. If a company's positioning, knowledge, relationships and judgement can live in a versioned, improvable body of skills, then the "soul" of the company stops being a slide deck plus a founder's memory — and becomes something you can address directly. The destination, in time, is that you will not operate a company's tools at all: you will talk with the company itself — a sentient enterprise that knows what it knows and reasons in its own voice, while a human still sets the intent and makes the final call.

That is the bet a small but serious group is now making. Projects like Incarnation.AI describe a "virtual incarnation" of a person or an organisation — an entity that embodies their character and values, reasons in their stead, and can even sit in meetings on their behalf. Others, like Soulsys, are building the memory-and-identity layer that gives an agent a coherent self over time; still others, like IncAgent, push the literal "one agent for the whole company" idea. Different bets, same intuition: the durable asset is the soul, and the natural interface to it is conversation.

SIFU is the down-to-earth, working version of that intuition — for one hemisphere of one company, go-to-market, with a human firmly in the seat. I find the destination genuinely exciting, and I am equally clear that the honest path there runs through evidence and human judgement, not around them.

One person I am watching closely on exactly this is Antoine Clerget, an investor in and advisor to EDAMAME, who is building Padanet (mens aucta — the augmented mind). Padanet goes after the human side of the same shift: as AI rewrites how work gets done week to week, it gives people and organisations a continuous, evidence-based view of how skills are actually evolving — signals instead of snapshots, evidence instead of declarations. It rhymes with everything above, and with what we sell. Keep an eye on it; I think there is important work coming.

Why a vendor's own tooling is worth writing about

Here is the throughline, and the reason this is more than a behind-the-scenes curiosity. A company that can observe itself, verify that what it did matched what it intended, and improve from the evidence is running exactly the playbook we bring to coding agents with EDAMAME: independently observe what a system actually does, compare it to what it was told to do, and turn the difference into evidence you can act on.

We sell verifiable, self-improving systems. It only seemed honest to build one to run the company — and to keep a human in the seat while we do.

Curious what that looks like for your coding agents and developer endpoints? See how EDAMAME secures the agent surface.