Step 1 · Mental model · Harness thesis ENPT
Alembic · Blind Planning Visual Course

Harness is the product

By the end you'll know why Agent = Model + Harness, why models commoditize, and why Alembic is a conductor — not a fourth generic orchestrator.

Plain version first — open technical layers per section.
1

The big idea


The blind planning prompt opens with an anchor thesis from the harness corpus: if you're not the model, you're the harness. TerminalBench leaders don't win on bigger prompts — they win on thinner, better-engineered harnesses. Pi tops benchmarks with four tools and 200–1000 token prompts. LangChain showed the same model jumping rank 30→5 by changing harness alone.

An agent is not a chat model. It is Model + Harness: the loop, tools, state, guardrails, verification, permissions, and orchestration that turn inference into reliable work. Prompt engineering is what you ask; context engineering is what fits in the window; harness engineering is everything else that makes the system operate.

Alembic is the product name for that harness layer in the Appfy stack — repo appfy/alembic, packages @alembic/*, state at ~/.alembic/runs/<id>/. It is deliberately not a fourth copy of Droid/Claude/Kimi UIs. It is the conductor that fuses three operational primitives (Mission, Workflow, Swarm) under Loop Engineering gates.

Think of it like… an orchestra. The LLM is one instrument section — powerful, interchangeable. The harness is the conductor, score, rehearsal rules, and recording booth. Alembic holds the baton: it does not play every note, but nothing reliable happens without it.

Corpus anchor (PROMPT-BLIND-PLANNING § Corpus)

Source: synthesis-harness-distribution/source.md embedded in PROMPT-CORPUS-EMBEDS.md. Key claims the blind agent must operationalize, not parrot: Cloud Code leak ~80% value in deterministic harness code; Terminus top-three with 1-keystroke harness; Dan (yt-156): "Without the agent harness there are no agents".

12 harness components (LangChain Anatomy embed)

Orchestration loop · Tools + scoping · Memory · Context · State · Errors · Guardrails · Verification · Permissions · Subagents · Observability · Co-evolution. Alembic plans must map each to a concrete runtime verb — not a bibliography.

Taxonomy (do not conflate)

LayerWhat it isAlembic role
Prompt engineeringWhat to askUnit prompts, forge grill
Context engineeringWindow fuelCompaction, embeds injection
Harness engineeringHow system operates@alembic/harness core
Loop engineeringMeta control planePrelude + Coda gates
2

In one picture


Model (LLM) + Harness = Agent Alembic Conductor Mission · Workflow · Swarm + Loop Engineering
Agent = Model + Harness. Alembic sits on the harness layer as conductor — extracting DNA from three products without cloning their UIs. Source: PROMPT-BLIND-PLANNING.md § O produto: Alembic.
3

In the code


Alembic's existing spine — what blind plans must extend, not greenfield.

@alembic/harness — Conductor (transport-neutral)
# Already implemented spine
HarnessCore.start | fanout | poll | report
EventBus + swimlanes
Council NO_GO → abort fanout (fail-closed)
CLI · HTTP+SSE · MCP read-only
loop-engineering/SKILL.md — control plane
LEARN → ANALYZE → EXECUTE ONE → VERIFY → DECIDE
Gates: Scope | Proof | Course | Publish
Validator ≠ builder

Product done-when (from blind prompt)

alembic run --goal GOAL.md --yes

≥1 Mission-sequential segment
≥1 Workflow-script (alembic.plan.ts)
≥1 Swarm-parallel (items[] + ramp)

Council GO · Proof Gate exit 0 · Validator ≠ builder
Artifacts: ~/.alembic/runs/<id>/events.jsonl, wire.jsonl, review.md
4

Try it: three layers of value


Click each layer. See where commoditization stops and harness investment begins.

Commoditizes: raw reasoning, token generation, benchmark scores shift quarterly.
Not enough alone: no persistent state, no proof gates, no programmatic runner.
Where value compounds: tools, state on disk, verification loops, permissions, observability.
Corpus proof: same model rank 30→5 (LangChain embed in corpus-cache).
Conductor fuses: Mission DNA + Workflow DNA + Swarm DNA + Loop Engineering.
Not: copying /missions, /workflows, /swarm UIs.
Next: evidence discipline — how blind agents must prove every claim before planning Alembic.