ensembly

Product direction

What the layer is for — not a shipped demo. Proving it needs costly real workflow evals.

One writer. On the operator machine, a single kernel holds the ops database and writes live status, gates, and irreversible acts. Grok Bot and other clients read and propose; they do not each hold canonical truth.

Gates

Agents and platforms capture context and draft output. Irreversible acts wait on approve/deny gates: send, pay, merge, submit. Plain gates — no insider command dumps.

ensembly holds live status and pending gates between sessions. Complement the bot loop; do not compete with it.

Hard problems

Three problems chat-scale products rarely productize — and that the pillars above must earn through evals:

• Ledger time travel — what was true at a past moment, from the ledger, not chat.
• Slice flight — work, gates, places, and clients as a cut through time, not a static repo map.
• Off-map location — private place-and-time facts that do not belong on public stack maps.

Frontier vs daily tools

Frontier models earn their keep: higher intelligence, steeper learning curve, real dollars and tokens. Use them where the work demands it.

Free tools and open-weight models handle daily assist. A frontier can hand work down when the pattern is clear.

Humans scatter tools, data, and effort across ecosystems. No harness controls what lives outside its boundary. CRDT-style sync everywhere is not the realistic story.

Sync

CRDT — Conflict-free Replicated Data Type: offline copies merge without a central lock. The paper assumes controlled replicas and a clean update stream.

Real scatter — chat threads, files, agent wire protocols and cloud sync, skewed clocks — can merge conflict-free on paper and still be wrong for gates, approve, and deny.

Multi-master CRDT is not solved for operator truth. The honest path: ledger, careful sync, one kernel writer on the operator machine when live status matters.

Hard gap

The remaining work needs real workflow evals — costly, which means capital. A frontier lab may buy ensembly later. This is product truth, not a loud pitch deck.

Evals on verified artifacts: correct, effective, efficient. Generational gain is harness and evals — not daily kernel churn.

Palantir vs ensembly

Palantir Foundry, AIP, and Ontology cover governed enterprise data, HITL actions, and agents that stage proposals. ensembly's wedge is different: coherent state under personal multi-harness scatter — chat never owns the ledger. Cousins in HITL thinking; different customer and tenancy.

Related: Sovereign AI, Hybrid Routing, Token Optimizzing (Jul 2026) — hybrid routing, token discipline, Palantir Ontology for governed enterprise context.

Aligned incentives

One writer on the operator machine holds the ledger, live status, and pending gates. Bots read and propose; they do not each hold canonical truth. Less re-briefing, fewer stalls when state was wrong or missing — that is the job.

Frontier labs earn when outcomes match intelligence: operators stay in flow instead of stuck on bad state. Stalls burn tokens without outcomes; flow spends tokens on finished work. Cheaper frontier pricing may help affordance — hope, not a promise or partner deal.

Free and open-weight tools handle daily and delegable work — cost-free or cheap enough to run locally. A frontier hands down when the pattern is clear; the kernel keeps status either way.

Operators get both: frontier depth where it matters, cheap daily cover elsewhere — one coherent, correct pulse across data, process, and place. The usual path is hundreds of process tools, a poorly adopted AI stack, failed applied AI, and everyone partly unhappy.

Without buy-in

Same as any product: competitors ship similar coordinators. That is fine. The choke — scattered tools, chat memory as truth, failed applied AI — still gets addressed by ensembly-like products. Category formation, not monopoly.

Process tools

Process-tool makers either talk to ensembly, or the harness bridges them. Tools read and propose; one writer on the operator machine holds the ledger, live status, and gates. Coherent state stays on the kernel — not in each app's private map.

Then process tools become flavors — visualization and UI flows across markets and geography, like people buy different cars and phones. Taste, not source of truth.

Kernel law: HITL / HOOTL

Automate the digital. Surface the physical. Wait only for permission. Make the truth playable.

HITL — human in the loop: deliberate verification and steering. HOOTL — human out of the loop: swarm clears digital thrash; physical pickups and irreversible gates wait for permission. Auth and physical gates never self-approve.

• Fitness for Human in the loop & out of the loop
• Pulse instead of dump
• Archive is not memory
• HITL / HOOTL
• EEaaS to agents

Use cases

Ask