- Project
- SuperHarness
- Looikos cluster
- Infrastructure & Agent Platforms (the harness layer itself)
- One-line
- The brand under which Andy teaches harness engineering, releases stress-tested open-source Hermes-based harness plugins, and sells multi-five-figure harness buildouts. The renamed, recentered successor to Symphony AGI, anchored on the domain name Andy secured.
- Status
- Active brand direction as of 2026-07-03. SUPERSEDES as the brand identity; that deck's market read, valuation comps, personas, and feature-factory analysis remain in force and are referenced here, not repeated. The Hermes/Harness V2 build continues under this name.
1. What it is (the one-paragraph truth)
SuperHarness is the harness engineering brand of the Looikos ecosystem: the public identity under which Andy teaches the discipline, releases stress-tested open-source plugins for the Hermes scaffold, and sells multi-five-figure harness buildouts to organizations whose agentic ambitions have outgrown their duct tape. A harness, in this vocabulary, is everything around the model that turns raw intelligence into dependable work. The brand's central claim is anatomical: the industry is still arguing definitions and semantics, but every serious harness resolves into the same set of organs (prompts, skills, personas, model selection, hooks, reference material, templates, agent SOPs, data models, integrations, memory, evaluation, observability, continuous integration, workflows, loops, error handling, navigation, operations), and the engineering practice that treats them as one body is what separates a demo from an operation. The name is new as of 2026-07-03 and anchored on the domain Andy secured. The substance beneath it is the live Harness V2 build the Symphony AGI deck documented, recentered from "orchestration platform" toward "the discipline, taught and sold." Symphony AGI's conducting philosophy survives intact inside it: you conduct a swarm by understanding what the piece is for, not by adding players.
Andy's framing (2026-07-03 directives, condensed with fidelity): SuperHarness is "a Hermes-based harness" brand. The process he designed and is building for himself is powered by OpenAI's Codex as the most powerful coding model, with Claude Code as the orchestrator and personal-assistant layer (driven through the desktop app or through Warp in the terminal), and open source carrying the heavy lifting and parallel execution when multiple perspectives are wanted. A harness breaks down into coherent parts with specialized systems in each: the prompts (readmes, "a ton of different markdown files"), skills, personas, model understanding and testing, hooks, reference materials and examples, Jinja templates and the process for managing and updating them, guides and SOPs for agents, data models grounded in Pydantic-as-IR (the discipline ScatterModel productizes), integrations across the agent protocols (MCP, A2A, ACP, ADK, GUIs, "there's so many of them"), memory (with WikiDesignCo as part of it), evaluation rubrics, observability, continuous integration pipelines and feature factories, workflows (orchestrations, LangGraph, DAG-thinking like Airflow), loops ("the modern version of prompts... a big part of what makes agents so powerful now"), errors and retries for reliability and durability, navigation (CLI, grep, APIs, "they need to navigate it somehow"), and operations (an operational ledger with the nine-stage hierarchy, "kind of like a CRM... built in directly to harnesses"). His conviction, in close paraphrase, is that it's so early people are still arguing about semantics, but all of these are going to be vital, "like the parts of a cell, the organs of a body." The brand builds notoriety by intermittently open-sourcing stress-tested pieces, and monetizes as teaching plus "multi five-figure harness buildouts."
Reading between the lines. Three moves sit inside the recentering.
First, the rebrand shifts the product from platform to practice. Symphony AGI was framed as the thing you run on; SuperHarness is framed as the thing you learn and hire. That's a sharper wedge into the same market the Symphony AGI deck mapped, because the buyer's trust problem (a hundred wrapper demos, a fear of science projects) is answered better by a named discipline with a public body of work than by another platform pitch. The platform stays on as the reference implementation the practice teaches from, which is the move the whole ecosystem runs: the working artifact is the demonstration.
Second, the three-part Hermes setup assigns model classes to seats instead of betting the brand on any one model. Codex takes the heavy-coder seat because it's the strongest pure implementer; Claude Code takes the orchestrator seat because holding context, dispatching, and grading is its home ground; open-source workers take the parallel fan-out seats because multiple cheap perspectives beat one expensive one on wide problems, with a stronger model synthesizing the winners. The scaffold is the constant and the seats are swappable, which is the same model-agnostic credit-quality argument the Symphony AGI deck made in its finance angle: the harness proven on the cheaper model upgrades in one line.
Third, the anatomy is the curriculum, and the curriculum is the moat. Each organ is a teachable module, a sellable buildout component, and an open-sourceable plugin surface, which means the brand's three revenue motions (teach, release, build) all decompose along the same lines. A competitor can copy any single released plugin; copying the assembled anatomy, with the tested judgments about how the organs compose under production load, requires the operational corpus, and that corpus lives in WikiDesignCo where it compounds privately.
3. The anatomy of a harness (the curriculum spine)
The list of organs is the brand's load-bearing content: Andy's decomposition, assembled from building these systems daily. Each organ is a real subsystem with its own engineering practice.
Prompts are the tissue everything grows from: readmes, operating contracts, the dozens of markdown files that define who an agent is and how it should think. They're versioned artifacts with owners and review discipline, not strings buried in code. Skills package procedural knowledge into loadable units, keeping the base prompt lean while making deep procedure available on demand. Personas shape tone and judgment per role, and they make the difference between five agents that all sound like the base model and a crew whose members can be trusted with different work. Model selection and testing treats the roster as a portfolio decision, verified by testing rather than vibes and revisited as the frontier moves; Hermes exists because no single model wins every seat.
Hooks are the mechanical enforcement layer: code that fires on tool use and writes, blocking the em dash, rejecting the invalid model ID, catching the deferral. Prose rules bind an agent that chooses to follow them; hooks bind everyone, and anything load-bearing eventually graduates from instruction to hook. Reference material is curated context, and getting the right reference into the right context at the right moment is a retrieval problem, one of several places the harness plugs into WikiDesignCo. Templates (Jinja, for everything that repeats) make output shapes deterministic where determinism is cheap, with template lifecycle managed as its own process because a stale template silently propagates a stale standard. Guides and SOPs for agents are operating procedures written for machine readers: how to hand off, how to report, what evidence a claim requires, with the nine-rung hierarchy, which runs from Mission to Event, as the master SOP.
Data models ground every structure in Pydantic as the intermediate representation, so every message and artifact an agent produces has a typed shape that survives translation across storage and transport; ScatterModel is the sister asset that productizes this discipline, one typed model projecting from a single IR to database, backend, frontend, and templates (see the scatter-model deck). Integrations answer how the harness connects to the world, and the answer is plural: MCP as today's workhorse, A2A for agent-to-agent traffic, ACP, ADK-style toolkits, plain APIs, CLIs, SDKs, sometimes a GUI driven directly. Knowing which protocol fits which seam is a core harness-engineering judgment. Memory is layered by time and truth, with WikiDesignCo as the deep layer: the corpus, the knowledge bases, and the metagraph agents query before acting and write back to after. A harness without a memory architecture repeats its mistakes with great confidence.
Evaluation rubrics define what good means before work begins, and in a loop-driven harness the evaluator is most of the system's real intelligence. Observability makes the harness auditable, LogFire spans on every operation of consequence, under the rule that what isn't in the traces didn't happen. Continuous integration keeps the harness itself improving: prompts, skills, and hooks flow through the same review discipline as code, because they're code, and the feature factory is this organ at domain scale. Workflows structure multi-step work as explicit graphs, LangGraph orchestrations with DAG-thinking a data engineer knows from Airflow; determinism lives in the workflow, judgment lives inside nodes.
Errors and retries are the durability organ: retry policies, idempotency, loud typed failure events, and never a fallback that masquerades as success. A harness earns trust exactly as fast as its failure handling deserves it. Navigation is how agents move through their world (a shell with grep, filesystem access, an API surface, a browser), and mismatched navigation is a silent capability killer. Operations closes the anatomy: the operational ledger that knows the task, the project, the owner, and the status, structured by the same nine rungs. The ledger is CRM-shaped machinery built for agents as first-class users, the territory PRD Master develops into a product (see the prd-master deck), because a harness without an operational ledger can't be managed, only watched.
4. Loops (the organ the era is named after)
Andy weights loops heaviest of all the organs, and the industry has just caught up to why. Loop engineering is the emerging discipline of designing the cycle that drives agents (act, observe, evaluate, decide, repeat) rather than prompting each step by hand, and the 2026 practitioner literature now names it as the successor layer to prompt engineering: the highest-leverage design surface has moved from the words you type to the cycle that keeps an agent productive, verified, and on-goal, including while you sleep.
Andy's version of the loop is specific, and it's the one this brand teaches. A fresh agent makes an attempt against pre-established criteria and publishes its work. A reviewer agent evaluates the attempt against those criteria. A third agent reads the review, the original, and the requirements, does its own research, and makes a better attempt. Each participant runs with a lean, purpose-sized context instead of one bloated window, which is why loops produce more quality per token than monolithic sessions: the loop does its action, completes, and keeps going, incrementally and iteratively improving. The external literature lands on the same structural conclusion the practice found first: the verifier is the bottleneck and the crown jewel, because the generator retries cheaply forever, so the quality of the evaluation criteria decides whether all that motion compounds into value. Loops are to this era what prompts were to 2023, and the evaluation-rubric organ is where a SuperHarness buildout earns its fee.
5. What carries over from Symphony AGI (referenced, not repeated)
The Symphony AGI deck (2026-06-21) remains the standing research for everything the rebrand doesn't change, so that research lives in one place. SuperHarness carries over in full, and cites from there, the conducting design philosophy, which derives the topology from the domain's intent, not from imposed volume; the three-angle valuation with its comps (Cognition, Imbue, Replit, Cursor, Zhipu, the MosaicML strategic-multiple logic) and the credit-quality argument for self-hosted, model-agnostic open-core; the artifact-marketplace third door, which treats every reusable unit of agent engineering as a versioned, cataloged asset, a position the incumbents won't occupy; the feature factory as the productized unit; and the five personas with their PST world model. Where that deck says Symphony AGI, read SuperHarness. The metagraph records the supersession as a bi-temporal edge, so the old belief stays queryable as history instead of being erased. The ecosystem treats every renamed belief that way.
The rebrand adds four deltas on top of that base. The brand's center of gravity is the practice (teach, release, build) rather than the platform; the anatomy is the public curriculum and the buildout checklist; loops are promoted from an implementation detail to a headline discipline; and the WikiDesignCo dependency is now a build in flight rather than a plan (see the wikidesignco deck), which converts the memory organ from aspiration to integration work.
6. Business model and the WikiDesignCo relationship
Revenue stacks in three tiers, and the tiers are the same three motions the anatomy decomposes into. The buildout tier sells bespoke harness engineering at multi-five-figure engagements: an organization with real workloads gets a Hermes-based harness designed around the full anatomy, with the evaluation rubrics, the observability, and the operational ledger that make it maintainable after the engagement ends, priced against the fear it resolves (the science project that never ships). The teaching tier packages the discipline itself, the anatomy as modules and loop engineering as the flagship, with Data Monastery as the neighboring teaching brand for the data layer beneath (see the data-monastery deck). The open-source tier is the top of the funnel rather than a revenue line: internally stress-tested Hermes plugins released deliberately and intermittently, each one a proof artifact aimed at the engineers who influence buildout purchases.
The WikiDesignCo relationship runs in both directions and is the deepest dependency in the ecosystem. Downward, every organ that touches knowledge runs on the platform: memory queries the corpus and the metagraph, reference material is retrieved from knowledge bases through their MCP connections, templates and rubrics live as versioned assets, and the operational ledger's records land as typed, provenance-bearing data. Upward, the platform is operated by harnesses, so every SuperHarness advance upgrades the crew that runs the corpus. Both share the staging the 2026-07-03 directives set out: the platform is the environment, humans and CLI agents are its users through the same surfaces, Gemini and Vertex carry the in-platform processing for now (LangGraph workflows or direct calls, Gemini being the model that reads video), and resident Hermes-class agents arrive later, low-level processing first, while real-time acting stays with humans and CLI agents. The ContentFactory live instance is the laboratory where that staging earns its keep on a paying customer today.
7. Priority read (feeds the value rubric)
SuperHarness inherits Symphony AGI's position on the dependency graph as the apparatus every brand is conducted by. It's foundational alongside WikiDesignCo, and the two form the substrate pair: the harness runs the work, and the platform remembers it. The work now is to continue the Hermes/Harness V2 build under the new name, and let the anatomy structure both the internal practice and the first public teaching material. Next, once WikiDesignCo's walking skeleton lands, the work is to wire the memory organ to real workspace knowledge bases via MCP, and cut the first open-source plugin release from a stress-tested internal piece. The thing to watch is the loop-engineering conversation, which is moving fast in public, so the teaching material's framing gets re-grounded against the literature at each authoring pass rather than assumed stable. What to leave alone is unchanged from the Symphony AGI read: never rebuild what the scaffold rents (model runtimes, commodity orchestration primitives).