Agent-stack consolidation vs the Modern Data Stack

Founder shared a tweet describing a two-person team's AI stack and asked: anything to evaluate, or over-engineered glue? He drew a parallel to the Modern Data Stack era (niche vendors → winner consolidation).

The stack in the tweet

Codex / Claude Code — "workbench"
Hermes agents — "workforce"
Orgo — Hermes agent infrastructure
Slack — communication layer
Second Brain — lives in GitHub
Composio — tool access
GHL (GoHighLevel) — CRM / sales automation

Their own headline: the Second Brain (the one piece they built rather than rented) is "the single most impactful thing we've built."

What the unfamiliar pieces actually are (verified 2026-07-05)

Hermes Agent — Nous Research's open-source agent: builds skills from experience, refines them in use, keeps a persistent user model across sessions. (GitHub, docs)
Orgo — cloud computers for AI agents: sandboxed Linux VMs provisioned by API call, sub-second boot, blast radius contained to the VM. Supports Claude computer use, OpenAI CUA, Hermes, custom harnesses. (orgo.ai, Orgo Hermes guide)
Composio — managed tool-access/auth layer (large SaaS connector catalog) agents call instead of hand-wiring each integration.
GHL — GoHighLevel, agency CRM. Vertical app, not stack-interesting.

Piece-by-piece mapping to what RDCO already runs

Their piece	RDCO equivalent	Delta
Claude Code workbench	Same (Ray's harness)	none
Hermes "workforce"	Subagents + brigade stations	Hermes's novel bits (skills-from-experience, persistent memory) = /skillify + /improve + vault memory, native
Orgo VMs	Owned Mac mini + SPCS deploy target	They rent isolation per-VM; we own the box
Slack comm layer	iMessage + Discord channels	none
Second Brain on GitHub	rdco-vault (+ QMD search, typed graph, gates)	Ours is deeper: typed edges, search, verification gates
Composio	Direct MCP wiring	The one real job in the list; MCP is commoditizing exactly that layer
Weekly ingest skill → Second Brain	Nightly crons, steward/cellar landing	Ours runs nightly with provenance discipline

Verdict (sent to founder)

Nothing to adopt. Not over-engineered glue either — it's a coherent rented copy of the stack we built native, right-sized for a two-person team that wants zero infra ownership. The tweet inadvertently validates the vault-first bet: the one component they built themselves is the one they call most impactful.

The MDS parallel holds, with one sharpening. In the MDS shakeout the consolidators were the platforms underneath (Snowflake/Databricks absorbing the glue). Here the consolidators are the harness vendors: Claude Code/Codex have already absorbed workbench + workforce (subagents) + tool access (MCP) + scheduling (cloud scheduled agents) natively. The durable layers are the ones you can't rent away: (a) the second brain — open-format files you own; (b) possibly sandbox infra (someone has to sell compute isolation). Everything in between sits in the squeeze path, Composio most visibly — a useful service whose moat is being standardized away by MCP.

Watch trigger (targeting-filter anchored): if CAF/client work ever needs dozens of SaaS integrations stood up fast (P4-scale client tool access), a Composio-style managed-auth layer earns a one-hour eval. Until that bottleneck materializes, it's an un-anchored shiny object. Orgo-style disposable VMs only become relevant if the brigade needs computer-use isolation beyond SPCS — future rail/runtime adapter candidate, not now.

Sanity Check candidate

Founder's MDS-consolidation frame is an original re-frame, not a restatement of the source ("the agent stack is speed-running the Modern Data Stack — and the consolidators this time are the harness vendors"). Flagged as a content candidate; his lived MDS experience is the differentiated angle.

Sources

Related: [[2026-06-04-agent-workflow-patterns-catalog]] · [[2026-06-13-open-knowledge-format-okf-assessment]] (OKF thread — same "portable second brain" territory)