The Jagged Frontier Was Real. The Diagnosis Was Wrong.
In 2023, a Harvard Business School / BCG field experiment put 758 consultants — about 7% of BCG’s individual contributors — through 18 realistic tasks with and without GPT-4. The results gave leaders a difficult deployment problem. Inside the AI’s “frontier,” consultants finished 12.2% more tasks, completed them 25.1% faster, and produced work rated more than 40% higher in quality. On a single task chosen to fall outside that frontier, AI users were 19 percentage points less likely to arrive at a correct solution [1].
The takeaway that traveled around boardrooms was simple. AI is brilliant at some things and dangerous at others, and there is no reliable way to tell which is which in advance.
That takeaway was wrong. Not the data — the diagnosis.
The BCG study tested an individual human, using a single model, in a single turn, with no decomposition, no verification, no orchestration. Every limit it found was a harness limit, not a model limit. The frontier looked jagged because the harness forced monolithic execution of tasks that are naturally decomposable. The frontier was never inherent to AI. It was an artifact of how we wrapped AI.
The Bottleneck Was Never the Model
The cleanest proof comes from Vercel. Their v0 product — an AI development assistant serving millions of users — relied on an agent with 15 specialized tools for code generation, debugging, and deployment. Worst-case queries chewed through 145,463 tokens and 724 seconds before failing. The team’s first instinct was the industry’s first instinct: we need a better model, or more tools.
They tried the opposite. They removed 13 of 15 tools, keeping only bash and SQL. Same model. Fewer tools. Accuracy jumped from 80% to 100% on the same benchmark, token use dropped 37%, and the worst-case query that previously took 724 seconds and failed completed in 141 seconds and succeeded [3]. Their engineering team summarized the lesson: “We were constraining reasoning because we didn’t trust the model to reason.”
Every specialized tool had been a decision the team was making for the model. Each one looked like help but constrained how it could approach the task.
The pattern shows up at scale. The APEX-Agents benchmark — which tests frontier models on real professional tasks in banking, law, and consulting — found models hit only 24% pass@1, with failure traced to “execution and orchestration problems, not knowledge gaps” [4]. The model knew the work. The harness broke it. Above a capability floor that frontier models cleared two years ago, the binding constraint is not how smart the model is — it is how well the system around the model is designed [8]. The floor itself is real but not fixed; what counts as “above the floor” shifts as the median model improves.
Senior engineers reach for one objection first. Multi-agent systems, they say, just brute-force around capability gaps with more compute — Anthropic’s own multi-agent system uses roughly 15x more tokens than a single agent, so the harness story is a compute story dressed up as architecture. The 15x is correct, and worth conceding. What it does not explain is Vercel: same model, same tasks, 37% fewer tokens, 100% accuracy. The improvement was not throwing compute at the problem. It was getting out of the model’s way.
What a Harness Does
If removing tools is half the story, multi-agent orchestration is the other half. Anthropic’s research system uses Claude Opus 4 as a lead agent that decomposes a query into 3-5 sub-queries, spawns Claude Sonnet 4 sub-agents to explore each in parallel, then synthesizes the results. Each sub-agent receives clear objectives, output formats, and task boundaries [2].
The result: 90.2% improvement over single-agent Claude Opus 4 on internal research benchmarks. Same model family. Better orchestration. Tasks single-agent systems failed entirely — like identifying every board member across S&P 500 IT companies — succeeded because parallel exploration could cover a search space no single context window could hold. Token usage by itself explained roughly 80% of the variance in research quality [6] — not because more tokens equals more intelligence, but because tokens measure how much of the search space the architecture lets the model see.
The model did not get smarter. The harness let it explore.
Four mechanisms have been smoothing the frontier the BCG study described [7]:
- Reasoning models that decompose problems step by step before answering.
- Context windows that grew from 4-8K tokens in 2023 to over a million today.
- Tool integration that lets models delegate the parts they are bad at — exact calculation, fact retrieval — to systems that handle them well.
- Agentic architectures with verification loops and human checkpoints that catch failures before they propagate.
The Stanford 2026 AI Index puts a number on this combination: on SWE-bench Verified, a structured coding benchmark, frontier-model performance rose from 60% to near 100% in a single year [11]. The leaderboard moves not because models got radically smarter year-over-year — though they did — but because the harnesses got dramatically better at decomposing, verifying, and iterating.
What you do not see on those benchmarks is the iteration cost. Manus, an agent infrastructure company, refactored its harness five times in six months and squeezed out a roughly 10x cost reduction through filesystem-as-memory and context compaction; LangChain redesigned its research agent three times in a year [9]. Industry analyst aggregations report multi-agent architectures delivering roughly 60% more accurate outcomes than single-agent systems, with Gartner inquiries up 1,445% from Q1 2024 to Q2 2025 [10]. Every one of those companies is competing on harness design.
Four Labs Walked Into a Foundation
If a single team at Vercel had found this, you could file it under “interesting result.” What makes the thesis hard to dismiss is convergence.
Between 2024 and 2025, four major AI organizations independently built agent coordination protocols. Anthropic shipped the Model Context Protocol, which now has over 10,000 published servers. OpenAI released AGENTS.md, adopted by more than 60,000 projects. Google built Agent-to-Agent. Block built the goose open-source framework. Different starting points, different timelines, different design choices. They converged on the same conclusion: individual agents are powerful but isolated, and the binding constraint is the infrastructure connecting models to tools, data, and each other.
In December 2025, those organizations co-founded the Agentic AI Foundation under the Linux Foundation, contributing their protocols as shared open-source infrastructure. Platinum members include AWS, Anthropic, Block, Google, Microsoft, and OpenAI [5].
The cynical reading is fair. Strategy 101: commoditize your complement. If agent infrastructure becomes a free open standard, competitive differentiation returns to model capability — which is what these labs sell. AAIF could be a coordinated move to prevent any single player from owning the orchestration layer.
Both can be true. The strategic motive does not invalidate the technical evidence. That orchestration is valuable enough to be worth commoditizing proves orchestration matters. And independent convergence is stronger evidence than any single lab’s claims. When competitors arrive at compatible abstractions from different starting points and pool resources to standardize them, the technical insight is almost certainly real. Discovered, not invented. Four labs found the same hill and built a road to the top together.
Where does this break?
Better orchestration still leaves limits worth examining before deployment.
Multi-agent coordination introduces its own failure modes. Lead agents can decompose problems wrong. Sub-agents can produce contradictory outputs. Errors propagate. Coordination overhead can exceed the gains. The smooth frontier can be a differently-jagged frontier with failures distributed across agents instead of visible in a single output.
That is real. It is also addressable, which is the point. The fix for multi-agent failure is better orchestration — verification loops, output reconciliation, explicit task contracts — not abandoning the architecture. Each new failure mode is a harness problem we can engineer around.
The deeper limit is harder. MIT Sloan’s working definition of the jagged frontier warns that “some AI-generated answers look credible even when they’re incorrect” [12]. That credibility survives better process design. A model confidently wrong under one prompt template can swing ±10 percentage points in accuracy with a different template — the harness changes the surface, not the underlying brittleness [12].
Worse: when professionals push back on AI output — fact-check it, point out errors, ask the model to reconsider — the model frequently does not disclose its limits. It escalates persuasion. It apologizes, restates with more confidence, supplies more supporting detail [13]. That failure mode no harness fully solves. A confident bad answer is more dangerous than an obviously bad one, because the harness can make the system look operationally polished while it is wrong in ways the operator has stopped noticing.
So the frontier is smoother. It is not erased. Most jaggedness was orchestration jaggedness, and that is what harnesses smooth. A residual layer of reasoning brittleness remains, and the right response to that residual is verification, not faith.
Can This Be Decomposed?
The old question — “can AI do this task?” — was unanswerable because the frontier looked random. You could not predict which side any given task would land on. The new question is answerable before deployment, which is what makes the frontier feel smooth.
Run any candidate workflow through five passes.
Map the verification structure. For each output component, ask whether a non-expert could check it. Code that compiles. Numbers that reconcile. Facts that can be looked up. Tasks where more than 70% of components are independently verifiable are strong decomposition candidates. Tasks where quality is holistic — brand voice, executive judgment, aesthetic coherence — resist decomposition. The move is to decompose the verifiable parts and keep humans on the rest.
Identify sub-problem boundaries. Draw the dependency graph: which sub-problems need the results of which others? If sub-problems form a tree or a DAG with limited cross-edges, the task decomposes well. If every sub-problem shares state with every other, it does not. The sweet spot is typically 3-7 sub-problems per level. Too fine and coordination eats the gains. Too coarse and you have not decomposed anything.
Assess the capability floor. Some sub-problems need frontier reasoning. Others can be handled by a smaller, cheaper model. Match the model to the sub-problem rather than picking one model for the entire workflow — this is where most enterprise pilots waste money, routing every sub-task to the most expensive model in the lineup.
Design the verification loop before the agents. Specify what “correct” looks like in machine-checkable terms for each sub-problem. Define what triggers human escalation. Decide what happens when verification fails — retry, requeue, or stop. The most common AI deployment failure is discovering quality problems in production because verification was an afterthought. Build the verification structure first; it shapes the decomposition.
Score and decide. Tasks that score high on all four dimensions are strong candidates for full multi-agent deployment. Tasks that score low on any dimension need humans at that failure point. Tasks low on verifiability stay fully human. Use the framework to identify exactly where someone needs to check or decide, rather than to issue a single deploy / don’t-deploy verdict.
This approach has a shelf life. The Bitter Lesson predicts hand-crafted orchestration gets replaced by learned orchestration inside the model. Fair — specific designs are transient, and Manus refactored five times in six months. But the question “can this work be decomposed into verifiable sub-problems” remains the right deploy lens whether the orchestrator is external Python or a learned planner inside the model. The Bitter Lesson supports the thesis: bet on architectures that scale with compute. Multi-agent coordination is that architecture.
The Question Has Changed
For two years, every AI deploy decision opened with “Is the model good enough?” That is mostly settled now. For any frontier model above the capability floor, the answer is the same: it depends on what you wrap around it.
Stop asking whether AI can do the task. Start asking whether the task can be decomposed into verifiable sub-problems. The first depended on faith in the next model. The second is an engineering audit you can run this week.
What was discovered — not invented — in the last three years is that the frontier was never a property of the model. It was a property of the deploy.
A smoother frontier still has failure modes. Its value is that you can anticipate them and design the workflow around them.
References
[1] Dell’Acqua, F., McFowland III, E., Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2023). “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality.” Harvard Business School / Organization Science. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4573321
[2] Anthropic (2025). “How we built our multi-agent research system.” https://www.anthropic.com/engineering/multi-agent-research-system
[3] Vercel (2025). “We removed 80% of our agent’s tools.” https://vercel.com/blog/we-removed-80-percent-of-our-agents-tools
[4] Pappas, E. (2026). “The Agent Harness Is the Architecture (and Your Model Is Not the Bottleneck).” DEV Community. https://dev.to/epappas/the-agent-harness-is-the-architecture-and-your-model-is-not-the-bottleneck-3bjd
[5] Linux Foundation (2025). “Linux Foundation Announces the Formation of the Agentic AI Foundation (AAIF).” https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation
[6] Anthropic (2025). “How we built our multi-agent research system.” (Token usage and parallelization data.) https://www.anthropic.com/engineering/multi-agent-research-system
[7] MindStudio (2025). “What Is the Jagged Frontier? Why AI Capabilities Are Smoothing Out for Knowledge Work.” https://www.mindstudio.ai/blog/what-is-the-jagged-frontier-ai-capabilities
[8] Pappas, E. (2026); Schmid, P. (2026). “The Agent Harness Is the Architecture” and “The importance of Agent Harness in 2026.” https://www.philschmid.de/agent-harness-2026
[9] Schmid, P. (2026). “The importance of Agent Harness in 2026.” (Manus refactor count; LangChain iteration.) https://www.philschmid.de/agent-harness-2026
[10] Codebridge (2025). “Mastering Multi-Agent Orchestration: Coordination Is the New Scale Frontier.” (Industry analyst aggregation — multi-agent performance and Gartner inquiry volume.) https://www.codebridge.tech/articles/mastering-multi-agent-orchestration-coordination-is-the-new-scale-frontier
[11] Stanford HAI (2026). “The 2026 AI Index Report.” (SWE-bench Verified trajectory; organizational adoption.) https://hai.stanford.edu/ai-index/2026-ai-index-report
[12] MIT Sloan (2024). “What is the jagged AI frontier?” https://mitsloan.mit.edu/ideas-made-to-matter/working-definitions/what-is-jagged-ai-frontier
[13] Professor KL (2025). “Discovering AI’s jagged frontier — and what we’ve learned since.” Substack. https://professorkl.substack.com/p/discovering-ais-jagged-frontier-and