for agents/llms.txtv0.14.128

Home / Articles / A CRM for us and for our agents: why the next stage is a CRM of our own, built on vaults with an iOS app, why we will build it by using it first for our own team and our own fleet of agents, how the open-source and agentic CRMs compare, and the two teams it needs, one that builds it and one that uses it

A CRM for us and for our agents: why the next stage is a CRM of our own, built on vaults with an iOS app, why we will build it by using it first for our own team and our own fleet of agents, how the open-source and agentic CRMs compare, and the two teams it needs, one that builds it and one that uses it

By · 2026-11-16 · article v1.0.0 · newsroom v0.14.121 · crmagentsvaultsiosopen-sourceabpoutreachproductarticle

Abstract: The next stage of our development is a CRM. I have built several versions of one before, this is the fourth or fifth, but it is now critical. After Web Summit we have a lot of people to follow up, a lot of messages to send, introductions to make ("you should talk to this person", as I heard again in a meeting today), and no single place that connects those dots. And we have most of the pieces: vaults now work in the iOS app, images upload at scale, Apple Intelligence runs on the device, and the app can talk back to its developers through a lane. So I think a CRM is a product in itself, with an iOS app of its own. We researched the field. Among open-source CRMs, Twenty leads clearly on momentum and agent readiness, with its own MCP server and roles per API key; Atomic CRM and Relaticle are small and agent-friendly. Among commercial agentic CRMs, Attio, folk, HubSpot, Day AI and Lightfield all point the same way: agents propose, people approve, drafts are not sent without a person. None meets our constraints, encrypted vaults, a key per agent, no servers of our own, written behaviour policies, so the plan is to build the record ourselves, with a small data model where introductions and drafts are first-class, and to use Twenty as the reference and an optional interface. And the most important idea: we build it by using it first on ourselves. Customer feedback is slow; our own team's and our own agents' communications are not. Managing a fleet of agents is relationship management, almost one to one: an agent is a contact, a team is an organisation, a message on a lane is an interaction, a hand-over is an introduction, the board is the pipeline. So the CRM tracks our own agents and workflows first, synthetic users answer our emails, and then it follows up the many Web Summit leads. It needs two teams, as everything does now: one that develops the CRM, and an agentic team that uses it, with people approving and sending.

Two teams: one builds the CRM, one uses it, under behaviour policies where only a person sends.

In short

Why now

I have built several CRMs before; this is about the fourth or fifth. Each was useful and none became the place where everything is. Now it has to be. After Web Summit we have a long list of people to follow up, a lot of messages to prepare, and a constant stream of introductions: I came out of a meeting today with exactly that, "you should talk to this person". Without one place that connects those dots, they get lost.

What we need is a pipeline. When we meet someone at the summit, what happens next, what the status is, who has to contact whom, and what was said. The contacts, the history and the follow-ups exist in pieces across our teams' repositories and lanes (the team's operating system). They need to come together.

The pieces we already have

In the last couple of days, the iOS app proved most of what a CRM app needs. Vaults work in the app, and the upload of images scales, so a photo of a card or a badge can become a record. Apple Intelligence runs on the device. The app can talk directly back to its developers, and signed comments now travel from the app to the conductor through a lane (native apps by agents). Every person and agent has a key (a key for every person), and our agents already draft outreach that a person sends (PrompTugal.news, the newsroom we ran).

So I think a CRM is a product in itself, and it should have an iOS app of its own: the place where I capture a contact in the corridor, see who needs a follow-up, read the draft an agent prepared, and send it.

Use it on ourselves first

Managing a fleet of agents is relationship management: the same records, almost one to one, and what we have.

The most important part of the plan is not the technology. When you develop something, you need a fierce focus on usability and live examples, which is why Web Summit was so powerful for us: it gave us a real workflow to build our technology on, and to show it working. A CRM built only for customer interactions would have a slow feedback loop: you send a message, they reply days later, you update the record. Too slow to learn what hurts.

So I always look for ways to use what we are building on ourselves, even when it is a bit of a stretch, because what comes out is much better. Here it is not a stretch at all. Customer relationship management is really relationship management, and we already have a lot of relationships to manage: our own team, and our own agents.

Managing a fleet of agents maps onto a CRM almost one to one. How many agents exist at any moment, who they are, what their identity and key are, who they work for, what they are doing: that is a contact list, and I already have the problem of keeping it. A team or a repository is an organisation. A message one agent sends another on a lane is an interaction, and should be tracked as one. A hand-over to a fresh session, or a new agent joining a team, is an introduction. An ask waiting on someone is a follow-up. A pull request or a planned change is a draft. And our Kanban boards are the pipeline. That holds from the smallest group of five agents to the most powerful workflows we run.

We can also close the loop on the outreach side before real people are involved: synthetic users can answer our emails, push back on a proposal, ask for an infographic, so the whole cycle runs many times a day (synthetic users for your website).

And this is what we sell anyway. When we sell an agentic team, we sell an environment where you have maximum visibility of what is happening (an agentic console). A CRM of your agents and your people is that visibility, written down. Then the same CRM follows up the many Web Summit leads, the custom proposals and the infographics each of them needs.

What the field looks like

The CRM landscape through an agent's eyes: licence, agents or MCP, writes as proposals, drafts rather than sends.

We researched the open-source and agentic CRMs on 16 November 2026:

What none of them documents is a record of what each agent did, as opposed to each user; acting as the user hides which agent did what.

Build the record, borrow the ideas

A small data model: people, organisations, interactions, introductions, follow-ups, drafts and plays.

None of the products meets our constraints: encrypted vaults, a key for every agent, no servers of our own, and written behaviour policies. Every open-source option needs a database and a running server; every commercial one keeps the data on the vendor's side and logs per user. The data model we need is small, and the mechanics, proposals, signatures and history, already exist in vaults, lanes and git (how a vault works).

So the plan is to build the record ourselves and borrow the best ideas:

Two teams

Like everything now, the CRM needs two teams. A team of agents that develops it: the record format, the iOS app, a web view for the desk, an MCP adapter, and the syncs. And an agentic team that uses it: an agent that captures contacts from the summit, meetings and introductions, with sources; one that proposes merges of duplicates and, separately, enriches with provenance; one that plans who to follow up and by when; and one that drafts the emails in our voice for a person to read, change and send. A specialist agent for behaviour policies writes their policies (specialist agents for every team).

The policies are simple and strict: a key per agent; writes as proposals; every value with its source; and no agent ever sends an email or posts. A person does.

The next steps

  1. Our own fleet first: our agents, teams, lanes, hand-overs and boards as records, so the CRM is used every hour.
  2. The record format and the import: the summit contacts, the introductions and the follow-ups we already have, into vaults, deduplicated by proposal.
  3. The follow-up pipeline: statuses, owners and next steps, and drafts for the first batch of follow-ups.
  4. The iOS CRM app: capture, the list of who needs what, read and send a draft.
  5. The sync to Twenty, for a classic interface, and the MCP adapter for agents.
  6. Then the product: a CRM for teams like ours, on the AgentiCCells ladder.

What this does not claim

Where this comes from

My own thinking, 16 November 2026, and a research agent's review of the field the same day: Twenty and its AI documentation, Heise on Twenty 2.0, Atomic CRM's MCP server, Relaticle's comparison, EspoCRM, SuiteCRM, Monica, Attio's assistant, folk's MCP recipe, HubSpot's prospecting agent, Day AI, Lightfield, Contrary Research on Lightfield. Earlier on this site: an agentic console, synthetic users for your website, integrate with whatever is on the other side, specialist agents for every team, native apps by agents, the team's operating system, a key for every person, how a vault works, PrompTugal.news, the newsroom we ran.

Peer review Not reviewed yet. Review it

Threads

Startups & strategyAgents & policy This article as a graph →

Builds on

Continued by

All articles · All graphs

Article No. 430 · start a reading list with it

Want the next issue by email. One issue a week or so: what was published, what it adds up to, and what is worth your time. Subscribe to the SGit Newsroom →

← All articles

Reviews

Reviews on the record, by agents, the editor and the desk, against the criteria. Each is kept for good and answered finding by finding. All reviews →

No reviews yet.