Canonical thesis · Sorbet

Boring tools.

AI agents need structured, durable state — records, ownership, status, history, permissions — to do real work. The 'boring' platforms that already model the world this way are quietly becoming the agent-economy substrate. We bet a basket of them compounds 25–30% IRR May 2026 → May 2029.

Why now

The shift is already in motion.

Five concrete reasons the next three years compound into the records-owners, not the wrappers.

  • OpenAI Symphony · Linear

    The Symphony spec uses Linear boards as the canonical task surface for autonomous coding workflows. Tools with rich records are already the agent control plane.

  • Atlassian MCP

    Atlassian's MCP server made its installed base of work state machine-readable and machine-writable. Acquisition rumours are strategically plausible — owning the SoR is owning the substrate.

  • Incumbent agent SKUs

    Salesforce, ServiceNow, Workday, Oracle, SAP have shipped agent SKUs in the last 18 months that monetise records they already own — not new content they have to build.

  • UX → data quality

    Linear's UX-driven cleanliness becomes a data advantage. Tools people enjoy generate cleaner state, which makes them better substrates than over-customised Jira instances people game.

  • Greenfield repositioning

    Greenfield 'agent platforms' are repositioning as thin wrappers over these substrates because they cannot replicate the records, permissions, or workflow engines.

Five-question diagnostic

The pillars the signal classifier scores against.

A tool is likely to become agent infrastructure when it answers yes to all five. Pillar-positive news drives add; two-or-more pillar negatives drive exit.

P1

Records, not content

Does the platform centre on durable records (tickets, accounts, employees) — not unstructured documents or chat?

P2

State machine, not labels

Are state transitions explicit, named, and enforced — not just tags on free text?

P3

Ownership, not implication

Is 'who is responsible' a typed field with permissions — not guessed from the message?

P4

Structural verbs, not conversational actions

Are domain operations modelled as named API verbs (close, reassign, escalate) — not buried in prose?

P5

Queryable history, not visibility

Can an agent read every state transition, attachment, and decision back through time via a stable API or MCP?

Structural claim

Four shifts that compound the bet.

  1. 01

    Records-rich platforms (Jira, Linear, Salesforce, ServiceNow, ERPs, source control, HRIS, ITSM, identity, data warehouses) become the backbone of agent systems.

  2. 02

    Content/chat-rich tools (email, Slack, generic docs) become secondary context sources that need expensive wrappers to feed records-rich systems.

  3. 03

    The boring tools win. CRMs, service desks, ERPs, calendars, source control, HRIS, finance — this is where meaningful agentic work actually happens.

  4. 04

    Incumbent moat compounds. Salesforce, ServiceNow, Atlassian, Microsoft, SAP, Oracle, Workday already own the records. Greenfield platforms either acquire one or wrap one.

What would invalidate

The watch-list.

If two or more invalidators trigger, the daily scan generates exit proposals across the affected buckets and a portfolio-level rebalance toward leaders + the hyperscaler hedge.

Not

What this portfolio is not.

  • — A short-term trading strategy. Daily scans collect signal, not turnover.
  • — A diversified market portfolio. It is concentrated in SoR + adjacent infra by design.
  • — A robo-advisor. The chat declines personalised investment advice.
  • — A backtested fund. The clock starts on the actual seed date — no historical alpha is claimed.