Closed beta · a working demonstration

The thread, held.

Cairn is the operations system we built for ourselves — a personal memory store wired to a team of fourteen named agents. It's how James cut his working week at sheepCRM from five days to three, without anything slipping.

If "agent-augmented operations" reads like marketing, this is the working evidence. Have a look around. If you're thinking about what something like this could mean for your team, get in touch.

The roster · 14 agents on the books

Named, characterful, on rotation.

No "AI assistant." Each one has a job description, a schedule, accountabilities, and a status report. They are accountable to each other, the Chief of Staff, and ultimately to James.

Studio Ghibli–styled portraits. Each role has a one-page prompt, a config file, and a stats stone. Cross-cutting Chief of Staff sits front-centre — she runs the morning brief, queue management, and tactical reviews.

Recently · from the dossier

What the team's been finding.

A rolling sample of stones the team has written this week — security advisories spotted before they hit the news, all-clear days, market intel, and the calls that needed surfacing.

Security Analyst

Flask cookie CVE caught the day it landed.

CVE-2026-27205 — Vary: Cookie omitted on certain access patterns. Filed for the sheepCRM stack before any advisory hit our inbox.

Security Analyst

Ten repos audited. All clean.

Axios 1.14.1 + 0.30.4 RAT attributed to Sapphire Sleet. Every repo audited, two using axios at safe versions, lockfiles untouched during the attack window. Closed in one session.

Triage Engineer

443,000 errors. None of them ours.

MongoDB SLOW_QUERY noise from validate $cmd. Five consecutive triage runs proved it was Atlas-internal — zero application call sites. Muted. Queue noise dropped accordingly.

Financial Advisor

Pension review: still watching, nothing breached.

Gilt 10-year at 4.95%, escalation threshold 5.2%. MPC meeting tomorrow — hold-at-3.75% is consensus. Status amber, no action needed today.

Financial Advisor

One pension fund quietly drifting below its model. Surfaced.

Each holding cross-checked against its declared benchmark and target allocation. One fund is sitting under model across consecutive reviews — small drift, but persistent. Flagged for the next session before it compounds.

Research Analyst

Confirmed: sheepCRM listed and current on MemberWise.

March 2026 directory entry verified — listed alongside Beacon, Glue Up, Wild Apricot, Fonteva. Resolves the "directory presence is unevaluated risk" flag from a previous scan.

Research Analyst

Organic search drives 53% of traffic — and the rules just changed.

AI-driven search shifts SEO from keyword targeting toward semantic intent. E-E-A-T heavily weighted; professional bodies are structurally well-positioned. SEO is now a retention lever as much as acquisition.

Pathfinder

Three briefings to land a strategic call.

A core milestone stalled for four-plus weeks while peripheral work continued. Pathfinder flagged it three briefings in a row — the third escalation finally surfaced it for a decision.

Archivist

Five new papers on how agent memory should work.

ICLR 2026 sweep — MIRROR (consolidation), MemAgent (RL-based stream), MAGMA (multi-graph), EverMemOS (engram), and A-MemGuard (defence against memory poisoning).

The shape of it

Three primitives, one loop.

Stones

Every fact, decision, person, source, project — typed and addressable. They grow edges between each other. Sensitivity-tiered, audit-friendly, owned and inspectable.

Roles

Fourteen agents on rotating schedules — error triage every 90 minutes, research watchlist twice a week, finance health daily. Each one a single page of prompt and a small config.

Dossier

The UI is a single editorial column. No dashboards. No widgets. Decisions get a kicker, a Fraunces headline, options as picks, and a verdict bar that slides up when you commit.

01 operator at the helm

14 agents on rotation

5→3 days a week at sheepCRM

∞ stones, edges, and counterpoints

How a day goes

The loop, in four moves.

01

The team scans the world

Cloud Canary errors, Linear issues, calendars, customer support inboxes, market signals, the cairn corpus itself. Each role reads only what its job demands — no general-purpose snooping.
02

They write stones

Findings become typed memory: a thought, a decision request, a tension, a project update. Stones grow edges to other stones automatically — provenance is a graph, not a footnote.
03

Chief of Staff triages

She closes duplicates and noise, batches the rest, and surfaces a small handful of calls. Three roll-calls a day check that every team member is still operating per their schedule.
04

James decides

Approve, reject, defer, or annotate. Decisions become resolutions; resolutions become Linear tasks. Tomorrow the team picks up the next layer. The thread never drops.

Honest answers

The questions you were going to ask.

Can we get access?: Cairn is in closed beta. Right now there's one operator and a team of fourteen agents — that's the working demonstration. If you'd like to explore agent-augmented operations for your team, or want Croftsware to build something similar, get in touch.
What's actually changed in practice?: James's working week at sheepCRM dropped from five days to three with the team in the loop — error triage, financial health checks, customer intelligence, security audits, research watchlists, communications drafts, strategy briefings. Same operational footprint, two days back. The agents don't replace people; they hold the thread when there isn't one to spare.
How is this different from Claude, Claude Code, or other agent assistants?: Cairn isn't a generalist. The fourteen roles each have a one-page prompt, a fixed schedule, and a job description — none of them are asked to "help with anything." They scan the world they were assigned to, write structured stones into a typed knowledge graph, and leave the rest alone. The graph is what compounds: every stone has provenance via edges, sensitivity tiers, and an audit trail. Generalist chatbots don't accumulate this kind of memory; cairn does.
What about security?: Three things. (1) No agent has direct egress — outbound actions like email, calendar invites, or document publishing live in purpose-built Lambda sorties with no agent-callable interface. (2) Postgres row-level security enforces three sensitivity tiers (normal / sensitive / secret) at the database, not just at the tool layer. (3) The lethal-trifecta principle is structural — no single agent has read-access to untrusted content, write-access to private data, and external egress at the same time. Each agent's capabilities are explicit and minimal.

Talk to us

Curious what this could look like for your team?

We're a software studio with a fully-instrumented case study running in closed beta. Half-hour exploration calls are free; we'll show you around the dossier and talk through what would translate to your operation.

Closed beta · Croftsware team can here