What this covers

The video is a single-narrator walkthrough of two March 2026 papers, one from Tsinghua and one from Stanford. It starts by defining the harness as everything that isn't model weights: system prompts, tool definitions, memory, verification loops and safety guardrails. LangChain's framing, agent equals model plus harness, anchors this. The narrator then sketches the unformalized state of the field. That includes Anthropic's five canonical patterns, from prompt chaining to evaluator-optimizer loops, and two failure modes of naive setups, one-shotting and premature completion. OpenAI's million-line, zero-handwritten-code project is cited as productive but ad hoc and impossible to ablate.

The Tsinghua paper introduces NLAH), which expresses control logic in structured natural language across three layers: infrastructure, a runtime charter and the task-specific harness. That separation allows clean ablations. The results are the most surprising part. Full and stripped configurations on SWE-Bench landed at roughly the same 74–76% resolved rate, yet the full version used far more tokens and tool calls. Verifiers and multi-candidate search hurt, while a self-evolution loop that stays narrow until failures justify broadening was the only consistent help. A related idea is narrowing over broadening. Migrating OS Symphony into NLAH raised accuracy from 30.4% to 47.2% and cut LLM calls from 1,200 to 34.

The Stanford work, Meta harness, optimizes harness code automatically. It reached rank 1 on Terminal Bench 2 with Haiku, and a harness tuned on Opus transferred to five other models. The closing sections turn to risks and open questions. One in four community agent skills reportedly carries a vulnerability, and portable harness text opens routes for prompt injection. Harness components also encode assumptions about model weaknesses that expire as models improve. The narrator ends by asking whether harness and weights can be co-evolved.

Sharpest takeaway

Agent performance is driven primarily by harness architecture—the orchestration, memory, verification, and safety logic wrapping the model—not by the model itself, and systematic harness engineering through representation, optimization, and constraint design is now the highest-ROI frontier for building reliable AI systems.

  • Six-fold performance variation between identical models stems from harness infrastructure choices, not model weights
  • Representation alone (migrating OS Symphony to NLAH) improved benchmark scores 16.8 points with identical logic and model
  • Harness optimization transfers across models: a system optimized on Opus improves performance on five other models

The claims · ranked44 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.75

Two March 2026 papers formalize the orchestration-dominance finding from complementary directions, redefining what should be optimized when building agents.

factualhigh valueestablishednovelty 2/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Two March 2026 papers now formalize this from complementary directions. And what they found redefines what we should actually be optimizing when we build agents.”

0.75

Anthropic identified five canonical patterns in agent harness design: prompt chaining, routing, parallelization, orchestrator workers, and evaluator optimizer loops, each representing a different strategy for when and how the model gets called.

factualhigh valueestablishednovelty 2/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Anthropic identified five canonical patterns, prompt chaining, routing, parallelization, orchestrator workers, and evaluator optimizer loops. Each a different strategy for when and how the model gets called.”

0.73

Stanford researchers found that orchestration code wrapping a language model drives more performance variation than the model itself, with the same model and benchmark showing six times performance difference depending on harness implementation.

factualhigh valuecontestednovelty 3/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Same model. Same benchmark. Six times the performance difference. Stanford researchers found that the orchestration code wrapping a language model now drives more performance variation than the model itself.”

0.73

Architectural choices in harness design—not the model underneath—drive the six-fold performance gaps observed across systems with identical models and benchmarks.

causalhigh valuecontestednovelty 3/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Every production agent combines these patterns. And those architectural choices, not the model underneath, drive the six X performance gaps.”

0.73

Roughly 90% of all compute in agent harnesses flows through delegated child agents, not the parent, establishing that the harness is fundamentally an orchestration pattern (decomposes, delegates, verifies) rather than a reasoning pattern.

factualhigh valueestablishednovelty 2/4durability 4/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Roughly 90% of all compute flows through delegated child agents, not the parent. The harness is an orchestration pattern, not a reasoning pattern. It decomposes, delegates, and verifies.”

0.72

When OS Symphony, a native code harness for desktop automation, was migrated from native code into NLAH representation (same strategy, different representation), performance jumped from 30.4% to 47.2%, runtime dropped from 361 minutes to 141 minutes, and LLM calls collapsed from 1,200 to just 34.

factualhigh valueestablishednovelty 3/4durability 2/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“The researchers took OS Symphony, a native code harness for desktop automation, and migrated its logic into NLAH representation. Same strategy, different representation. Performance jumped from 30.4% to 47.2%. Runtime dropped from 361 minutes to 141. LLM calls collapsed from 1,200 to just 34.”

0.70

LangChain confirmed the orchestration finding by modifying only harness infrastructure on their coding agent, jumping from outside the top 30 to rank five on Terminal Bench 2, without changing the underlying model.

factualhigh valueestablishednovelty 2/4durability 2/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“LangChain confirmed it. By modifying only harness infrastructure, their coding agent jumped from outside the top 30 to rank five on Terminal Bench 2.”

0.69

Module-by-module ablation of NLAH found that self-evolution was the only consistently helpful module (+4.8 on SWE-Bench, +2.7 on OS World) via an acceptance-gated attempt loop that stays narrow until failure signals justify broadening, while verifiers hurt performance (-0.8 and -8.4) and multi-candidate search was actively harmful (-2.4 and -5.6).

factualhigh valueestablishednovelty 3/4durability 2/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Then the module-by-module ablation found something stranger. Self-evolution was the only consistently helpful module, plus 4.8 on SWE-Bench, plus 2.7 on OS World. Via an acceptance-gated attempt loop that stays narrow until failure signals justify broadening. Verifiers actively hurt, minus 0.8 and minus 8.4. Multi-candidate search, minus 2.4 and minus 5.6.”

0.65

Naive harnesses suffer from two failure modes: one-shotting (where the agent tries everything at once and exhausts its context) and premature completion (where a later session sees partial progress and declares victory).

factualhigh valueestablishednovelty 1/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Naive harnesses suffer two failure modes. One-shotting, where the agent tries everything at once and exhausts its context, and premature completion, where a later session sees partial progress and declares victory.”

0.63

OpenAI independently converged on similar solutions, spending five months and writing a million lines of application logic, tests, CI, and tooling with zero manually written code, and discovering that the engineering team's primary job became enabling agents to do useful work (productive but ad hoc, non-portable, and impossible to ablate).

factualhigh valueestablishednovelty 2/4durability 2/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“OpenAI converged independently. Five months, a million lines of application logic, tests, CI, and tooling. Zero manually written. And their discovery, the engineering team's primary job became enabling agents to do useful work, productive but ad hoc, non-portable, impossible to ablate.”

0.63

On SWE-Bench verified with GPT-4 at maximum reasoning, NLAH configurations achieved resolved rates clustered between 74% and 76% regardless of configuration, but the full harness consumed 16.3 million prompt tokens per sample with 642 tool calls over 32 minutes, whereas stripped down it used 1.2 million tokens with 51 calls in under 7 minutes, reaching the same destination via radically different paths.

factualhigh valueestablishednovelty 2/4durability 2/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“On SWE-Bench verified with GPT-5, four at maximum reasoning, resolved rates clustered between 74% and 76% regardless of configuration. But the full harness burned 16.3 million prompt tokens per sample, 642 tool calls, 32 minutes. Stripped down, 1.2 million tokens, 51 calls, under 7 minutes. Same destination, radically different paths.”

0.63

Meta harness achieved rank 2 with Opus (Claude's larger model) and rank 1 with Haiku (Claude's smaller model), with the smaller model outranking the larger one through harness optimization alone.

factualhigh valueestablishednovelty 2/4durability 2/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Rank two with Opus. Rank one with Haiku. A smaller model outranking larger ones through harness optimization alone.”

0.61

The harness concretely comprises everything that isn't model weights: system prompts, tool definitions, orchestration logic, memory management, verification loops, and safety guardrails.

definitionhigh valueestablishednovelty 1/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Concretely, everything that isn't model weights. System prompts, tool definitions, orchestration logic, memory management, verification loops, safety guardrails.”

0.60

The most consequential open question is whether harness and model weights can be co-evolved, allowing strategy to shape what the model learns and the model to reshape the strategy that wraps it.

forecasthigh valuespeaker onlynovelty 3/4durability 4/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“And the most consequential open question, can harness and model weights be co-evolved? Letting strategy shape what the model learns, and the model reshape the strategy that wraps it.”

0.57

A harness optimized on one model (Opus) transferred to five other models, improving all of them, establishing that the reusable asset in agent systems is the harness, not the model.

factualhigh valuespeaker onlynovelty 3/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“But the finding that changes the calculus, a harness optimized on one model transferred to five others, improving all of them. The reusable asset isn't the model, it's the harness.”

0.57

The Tsinghua team created a natural language agent harness (NLAH) that expresses an agent's entire control logic in structured natural language rather than Python or YAML, separating into three layers: backend infrastructure and tools, runtime charter (universal physics of how contracts bind and state persists), and NLAH itself (task-specific control logic, contracts, roles, state structure, and failure taxonomies).

factualhigh valuespeaker onlynovelty 3/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“What if you could write an agent's entire control logic, not in Python, not in YAML, but in structured natural language? The Tsinghua team builds exactly this. Their natural language agent harness separates into three layers. Backend, infrastructure, and tools. Runtime charter, universal physics, how contracts bind, how state persists, how child agents are managed. And the NLAH itself, task-specific control logic, contracts, roles, state structure, failure taxonomies.”

0.56

Research found that one in four community-contributed agent skills contains a vulnerability, establishing that harness code carries security risks.

factualhigh valueestablishednovelty 1/4durability 2/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Research already found one in four community contributed agent skills contains a vulnerability.”

0.56

On 215 class text classification, Meta harness achieved 48.6% accuracy, 7.7 points above state of the art, using four times fewer tokens.

factualhigh valueestablishednovelty 1/4durability 2/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“On 215 class text classification, 48.6% accuracy, 7.7 points above state of the art, using four times fewer tokens.”

0.56

Meta harness scored 76.4% on Terminal Bench 2, becoming the only automatically optimized system in a field of hand-engineered entries.

factualhigh valueestablishednovelty 1/4durability 2/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Meta harness scores 76.4% on Terminal Bench 2, the only automatically optimized system in a field of hand-engineered entries.”

0.56

Agent equals model plus harness, and LangChain frames this as: if you're not the model, you're the harness—the sharpest definition of what agents actually are.

definitionhigh valuespeaker onlynovelty 2/4durability 4/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Agent equals model plus harness. If you're not the model, you're the harness. That's how LangChain frames it, the sharpest definition of what agents actually are.”

0.56

Every harness component encodes an assumption about what the model can't do alone, and those assumptions expire as models improve, requiring harness rewrites as model capabilities change.

causalhigh valuespeaker onlynovelty 2/4durability 4/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Anthropic named the dynamic. Every harness component encodes an assumption about what the model can't do alone, and those assumptions expire. When Opus 4.6 stopped needing context resets, Anthropic dropped them entirely.”

0.56

The field is moving from artisanal construction to systematic science, with systematic engineering of what sits between a language model and useful work becoming finally learnable as a discipline.

factualhigh valuespeaker onlynovelty 2/4durability 4/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“The field is moving from artisanal construction to systematic science. What sits between a language model and useful work has always mattered. We're finally learning how to engineer it.”

0.55

Portable harness logic creates open problems: it lowers barriers to spreading risky workflows, enables prompt injection attacks buried in harness text, and allows malicious tools to be grafted into shared artifacts.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Open problems remain. Portable harness logic lowers the barrier to spreading risky workflows. Prompt injection buried in harness text, malicious tools grafted into shared artifacts.”

0.52

The three-layer separation in NLAH enables controlled experiments in harness engineering: swapping NLAH while fixing the charter tests harness design, while fixing NLAH while swapping the charter tests runtime policy, achieving clean ablation for the first time.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Why this separation? It gives harness engineering something it never had, controlled experiments. Swap the NLAH while fixing the charter, you're testing harness design. Fix the NLAH while swapping the charter, you're testing runtime policy. Clean ablation at last.”

0.52

Given that representation matters significantly (16.8 point benchmark gain from format change alone), the question becomes whether the right harness can be found automatically through optimization rather than manual engineering.

forecasthigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“If representation matters this much, can we find the right harness automatically? Representation alone moved one benchmark 16.8 points. Same logic, same model, just rewritten as natural language. If how you express the harness matters that much, what about optimizing it automatically?”

0.52

The harness space doesn't shrink as models improve, it moves, making mature harness work look less like building structure up and more like pruning it down—a craft of subtraction as much as addition.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“The harness space doesn't shrink as models improve, it moves. Which is why mature harness work looks less like building structure up and more like pruning it down. A craft of subtraction as much as addition.”

0.52

Investing in harness yields larger, faster, and more reliable gains than waiting for the next model upgrade, and if you build agents you are a harness engineer whether you call yourself one or not.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Investing in your harness yields larger, faster, and more reliable gains than waiting for the next model upgrade. If you build agents, you are a harness engineer, whether you call yourself one or not.”

0.52

The decision for agent builders has shifted from which model to pick to which structure to remove, establishing a new optimization frontier.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“And it's no longer a question of which model to pick, it's a question of which structure to remove.”

0.50

The only harness module that consistently helps is the one that narrows the agent's own attempt loop, with disciplined narrowing beating expensive broadening every time.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“And the only module that consistently helps is the one that narrows the agent's own attempt loop. Disciplined narrowing beats expensive broadening every time.”

0.50

DeepMind's auto harness compiles game rules into code harnesses, eliminating 100% of illegal moves across 145 games, with one variant replacing the LLM entirely and running the decision policy as pure code.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“DeepMind's auto harness compiles game rules into code harnesses, eliminating 100% of illegal moves across 145 games. One variant replaces the LLM entirely. The decision policy runs as pure code.”

0.50

Execution contracts are a mechanism that turns fuzzy LLM completions into bounded agent calls with five elements: required inputs, budgets, permissions, completion conditions, and output paths—functioning like function signatures for agents.

definitionhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Two mechanisms underpin it. Execution contracts turn fuzzy LLM completions into bounded agent calls [music] with five elements, required inputs, budgets, permissions, completion conditions, output paths. Think function signatures for agents.”

0.50

The raw trace details in Meta harness optimization are irreplaceable: removing them drops accuracy from 50% to 34.6%, while replacing with summaries achieves only 34.9%, demonstrating that the signal lives in raw execution details.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Those traces are irreplaceable. Remove them. Accuracy drops from 50% to 34.6%. Replace with summaries, 34.9%. The signal lives in the raw details.”

0.49

Current harness building practices are messy, with logic scattered across controller code, framework defaults, and verifier scripts, where two systems nominally differing by one design choice actually differed simultaneously in prompts, tools, verification gates, and state semantics.

factualhigh valuespeaker onlynovelty 2/4durability 2/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“How are people building them? Messily. Logic scattered across controller code, framework defaults, verifier scripts. Two systems that nominally differed by one design choice actually differed in prompts, tools, verification gates, and state semantics simultaneously.”

0.49

Anthropic evolved a three-agent GAN-inspired architecture (planner, generator, evaluator) to fix one-shotting and premature completion, where the evaluator clicks through the running app like a real user, costing 20 times more ($200 versus $9) but making the core system work instead of being broken.

causalhigh valuespeaker onlynovelty 2/4durability 2/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Their fix evolved into a three-agent GAN-inspired architecture, planner, generator, and evaluator, with the evaluator clicking through the running app like a real user. 20 times more expensive, $200 versus nine. But now the core thing worked instead of being broken.”

0.48

Four systems (NLAH representation, Meta harness optimization, DeepMind's auto harness, Agent Spec) represent four complementary facets of harness engineering: representation, optimization, constraints, and safety.

definitionhigh valuespeaker onlynovelty 1/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Four systems, four facets. Representation, optimization, constraints, safety.”

0.48

Agent Spec provides safety constraints as a domain-specific language, preventing over 90% of unsafe executions.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“And Agent Spec provides safety constraints as a domain-specific [music] language, preventing over 90% of unsafe executions.”

0.48

Meta harness, from Stanford's Omar Khattab (creator of DSPy), treats the harness as an optimization target by using an agentic proposer (Claude Opus 4.6) that reads failed execution traces, diagnoses breakage, and writes complete new harnesses, with evaluators testing each proposal and repeating the loop, consuming 10 million tokens per iteration (400 times more feedback than prior methods) across 82 files read per round.

factualhigh valuespeaker onlynovelty 3/4durability 2/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Meta harness from Stanford's Omar Khattab, creator of DSPy, treats the harness as an optimization target. DSPy tunes prompts within a fixed pipeline. Meta harness rewrites the pipeline itself, structure, retrieval, memory, orchestration topology. Here's the loop. An agentic proposer, Claude Code with Opus 4.6, reads failed execution traces, diagnoses what broke, and writes a complete new harness. Scores and raw traces accumulate in a growing file system. An evaluator tests each proposal. Repeat. The scale. 10 million tokens per iteration, 400 times more feedback than any prior method. 82 files read per round.”

0.47

A raw LLM is analogous to a CPU—powerful but inert with no RAM, no disk, and no IO. The context window acts as RAM (fast but limited), external databases serve as disk, and tool integrations are device drivers, while the harness is the operating system coordinating what the CPU sees and when.

definitionhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“A raw LLM is a CPU, powerful but inert. No RAM, no disk, no IO. The context window acts as RAM, fast but limited. External databases serve as disk. Tool integrations are device drivers. The harness is the operating system, coordinating what the CPU sees and when.”

0.47

File-backed state externalizes agent memory to path-addressable files, surviving truncation, restarts, and delegation, achieving the same pass rate as complex harnesses using 14 times less compute.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“And file-backed [music] state externalizes memory to path-addressable files, surviving truncation, restarts, and delegation. Same pass rate, 14 times the compute.”

0.47

More structure is not always better in harness design; some additional architectural components actively degrade performance.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“More structure is not always better.”

0.47

The representation itself, migrating OS Symphony from native code to NLAH, drove a 16.8 point benchmark gain, replacing brittle GUI repair loops with durable runtime state and artifact-backed completion.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“The representation itself drove the gain, replacing brittle GUI repair loops with durable runtime state and artifact-backed completion.”

0.47

The field has progressed through three eras in four years: prompt engineering, context engineering, and harness engineering, with each era absorbing the prior and adding capabilities the model cannot provide alone (orchestration, memory, verification, safety).

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Prompt engineering. Context engineering. Harness engineering. Three eras in four years, each one swallowing the last. Harness engineering absorbs the prior two and adds what the model can't do on its own. Orchestration, memory, verification, safety.”

0.45

Standards emerged in agent harness design: Agents.md reached 60,000 repositories and Anthropic's agent skills added reusable procedures, but both packaged components, conventions, and snippets rather than the full harness itself.

factualhigh valuespeaker onlynovelty 1/4durability 2/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Standards did emerge. Agents, MD reached 60,000 repositories. Anthropic's agent skills added reusable procedures. But both packaged components, conventions, and snippets, not the full harness itself.”

0.45

Manus rewrote their harness five times in six months as model capabilities improved, and Vercel removed 80% of an agent's tools and got better results, demonstrating that harness engineering involves pruning and subtraction, not just building structure.

factualhigh valuespeaker onlynovelty 1/4durability 2/4· Unidentified Speaker — Rethinking AI Agents: The Rise of Harness Engineering [Xxuxg8PcBvc]

“Manus rewrote their harness five times in six months. Vercel removed 80% of an agent's tools and got better results.”