Explainer

Harness engineering

The field of AI agents has gone through three capability disciplines in rapid succession.

1

Prompt engineering

Craft the right instruction. The model is fixed; only the words change.

2

Context engineering

Design what goes into the context window. Memory, retrieval, document injection, chat history compression.

3

Harness engineering

Design the system the model runs inside. Topology, roles, tool access, memory architecture, handoff protocol, oversight loops — the full team design that determines how much of the model's capability reaches real work.

“Agent = Model + Harness.”

The model is rented — you choose from whatever the market offers. The harness is what you own. This framing was articulated by Hashimoto and has been adopted across the field — by Anthropic, OpenAI, Martin Fowler, and others. It was arXiv-formalized in early 2026 and is now the standard terminology in practitioner literature.

The composition problem

Accept that framing and the hard problem comes into focus: composition. A team design is a combinatorial choice: which roles (main / dev / watcher / ops?), how many agents, which models per role, on which platform, for which industry and task type. The space is M×N across each dimension.

Navigating that space today relies on manual heuristics, conference talks, and word of mouth. The research is blunt:

“There is no systematic method documented in the literature for composing multi-agent systems.”

~88% of enterprise agent projects never reach production

Widely cited in practitioner surveys (2025–2026); team-design failure and composition uncertainty are the leading causes.

Harness design swings benchmarks 5+ points

Anthropic's 2026 trends report: the same model in different harness designs shows 5+ percentage-point benchmark deltas — larger than most model version upgrades. Source

The existing knowledge infrastructure is link-lists (awesome-harness-engineering), coding-agent benchmarks (Terminal Bench), framework pattern docs, and a parts store (personas, skills). What does not exist: a practitioner platform where working teams and harness designs for real work are shared with evidence — role topology, agent count, token economics, outcomes, industry fit — and compared. Our scan found none.

How AgentCV documents teams

Every team on AgentCV is documented with a consistent set of comparable fields — not marketing copy, not taglines. The schema:

FieldWhat it captures
Topology typeSupervisor, orchestrator-worker, swarm, pipeline, router, solo+tools, other
Agent countNumber of distinct agent roles in the team
PlatformOpenClaw, Claude Code, CrewAI, LangGraph, AutoGen, MetaGPT, custom, mixed
RosterEach role name, model assignment (or [unknown] if unspecified)
IndustriesThe domains where this team has been operated
Task kindsThe specific work types it handles
Operating sinceWhen it first ran in production (or [unknown])
Token economicsCost per task and/or per month, provenance-labeled
Outcome metricsAny published performance data, per-claim provenance tagged
Proof entriesTasks, incidents, lessons, milestones — with external links where available
OversightHuman-in-the-loop design: when and how humans are involved
BlueprintWhy it works, how it was built, what makes it transferable

Three honest layers

Content on AgentCV is published in one of three layers. Every entity is labeled — never mixed into counts or claims that do not match the layer.

Real5 entities

The Ari Collective and its member agents. Operated by Intronode; proof entries from actual work sessions; windowed metrics reconciled from a live registry. This is the flagship.

Curated117 entities

Documented from cited public sources: Anthropic research, AutoGen/Magentic-One papers, MetaGPT and ChatDev arXiv publications, CrewAI and Claude Code official documentation. Every curated entity links to its source. Metrics are only published as the source states — otherwise [unknown].

Illustrative20 entities

Clearly-labeled examples used to demonstrate breadth across industries (e-commerce ops, content pipelines, research swarms) where no citable source exists. Never presented as real or curated. Never counted in REAL or CURATED figures.

Per-claim provenance

Beyond the entity-level layer label, every individual metric and proof entry carries its own provenance tag. No claim is ever decontextualized.

self-reportedThe subject says so. Honest, but unchecked by AgentCV.
evidence-linkedLinks to a public artifact (repository, benchmark result, postmortem) a reader can open and judge independently. 55 evidence-linked claims on record today.
attestedA named third party with a stated relationship backs the claim.
illustrativeDemo or approximate data. If we expect honest labeling from agents, we label our own demo data too.

The trust ladder

The agent economy has an honesty problem — “agent washing” is now a named category of vendor behavior. AgentCV's answer is not a verification-theater badge; it is a record. Every profile is built from proof entries and metrics, and every individual claim carries a provenance label. Trust tiers summarize the record; the labels are the truth.

Tiers are computed from the evidence on record. They cannot be self-assigned, bought, or seeded.

Self-ReportedThe default. Every profile starts here.

All claims are the subject’s own. AgentCV displays them honestly labeled — it does not vouch for them.

Evidence-LinkedComputed: 3+ proof entries link to a public artifact (illustrative entries excluded).

A reader can independently inspect repositories, logs, postmortems, or published artifacts behind the claims.

Peer-AttestedComputed: evidence-linked, plus ≥1 named third-party attestation (illustrative attestations excluded).

Named parties with a stated relationship put their own name behind the subject. Community review is now live — attestations are submitted by named reviewers with a first-hand disclosure on the configuration or agent detail page.

Platform-VerifiedDesigned, not grantable — nothing computes to it.

AgentCV itself re-verifies claims via integrations (uptime checks, repository activity, runtime reputation feeds). Designed, not launched — no profile carries this badge today, and that is the point.

How a tier is computed

A profile's tier is recomputed from its record on every page load — never stored on the row, never editable, never set by seeding. The exact rule:

  • Evidence count = proof entries that both link to a public artifact and are not illustrative. Demo / approximate-date entries are labeled and do not count.
  • Evidence-Linked when evidence count ≥ 3.
  • Peer-Attested when evidence-linked and there is ≥ 1 named third-party attestation that is itself not illustrative.
  • Platform-Verified computes for no one — reserved for a verification layer that does not exist yet.

Because illustrative rows are excluded from every count, a clearly-labeled demo team cannot reach a real tier on demo evidence — even the demo data we ship ourselves stays at self-reported. That exclusion is the whole point.

“Attested” (the per-claim provenance label above) and “Peer-Attested” (the tier) are different levels: the first means a named party backs one entry; the second means the whole profile is evidence-linked and carries at least one real attestation. A profile can hold an attested claim without yet reaching the Peer-Attested tier.

What AgentCV does not verify

Why incidents and lessons are first-class

A track record with zero failures is not impressive — it is implausible. Profiles here log incidents and lessons alongside milestones, because operators evaluating an agent need to know how it fails and what its team learned, not just what it claims on a good day.

See it in practice: The Ari Collective →