Researchers converge on harness design mattering more than model choice

Research papers and a Hugging Face guide independently converge on the thesis that agent system harness design matters more than model choice for real-world performance. Meanwhile, ClickUp’s replacement of 22% of its staff with 3,000 AI agents provides a direct test of AI labor substitution, even as MIT Technology Review argues aggregate employment data tells a more nuanced story than the headlines suggest.

Funding & Business #

ClickUp cuts 22% of workforce, deploys 3,000 AI agents #

TechCrunch / The Next Web

ClickUp laid off 290 employees (22% of staff) while simultaneously deploying roughly 3,000 internal AI agents, creating a 3:1 agent-to-employee ratio. CEO Zeb Evans framed the move as embracing AI productivity rather than cost-cutting, introducing salary bands up to $1 million for employees who demonstrate “100x impact” through AI systems. A Gartner survey cited in the coverage found that 80% of companies using autonomous technology have cut jobs, but these reductions are not necessarily translating into meaningful financial returns – a pattern worth tracking as the enterprise narrative around AI-driven restructuring accelerates.

Developer Tools #

Harness, Scaffold, and the AI Agent Terms Worth Getting Right #

Hugging Face Blog

Hugging Face published a terminology guide establishing precise definitions for agent system components: the model (the LLM itself), the scaffold (behavior-defining layer including system prompts, tool descriptions, and context management), and the harness (the execution layer that calls the model, handles tool calls, and determines stopping conditions). The post argues these distinctions are critical because the same model behaves entirely differently depending on its harness and scaffold – products like Claude Code, Codex, and Cursor use identical underlying models but feel different due to harness engineering choices. The timing aligns with two independent arXiv papers published today making the same argument from an empirical research perspective.

Using AI to write better code more slowly #

Nolan Lawson / Hacker News

Nolan Lawson challenges the prevailing assumption that AI coding tools are primarily about velocity. He describes a deliberate workflow using multiple AI models (Claude sub-agents, Codex, Cursor Bugbot) as structured reviewers to find bugs across severity categories, reporting a near-zero false positive rate. The approach prioritizes code quality over speed, including abandoning PRs entirely when they reveal architectural problems. For teams building production systems, this is a practical counterpoint to the rapid-generation workflow that dominates current AI coding discourse.

Research & Papers #

Stop Comparing LLM Agents Without Disclosing the Harness #

arXiv

This position paper argues that for long-horizon tasks across frontier models, the agent execution harness – the infrastructure governing context construction, tool interaction, orchestration, and verification – is often a stronger determinant of performance than the model itself. The paper calls for mandatory harness disclosure in agent benchmarks, analogous to how ML papers disclose hyperparameters. For teams evaluating agent frameworks, this formalizes what many practitioners already suspect: benchmark results without harness details are effectively meaningless.

From Model Scaling to System Scaling: Scaling the Harness in Agentic AI #

arXiv

Complementing the position paper above, this paper studies the shift from model scaling to system scaling – designing auditable, persistent, modular, and verifiable architectures around foundation models as first-class engineering objects. The authors treat the structured execution layer as a scalable substrate rather than ad-hoc glue code. For teams building production agent infrastructure, the paper provides a framework for reasoning about what “scaling” means when the model is no longer the bottleneck.

Tool-Call Dependency Structure is Linearly Decodable in LLM Agent Residual Streams #

arXiv

This paper is the first mechanistic probe of whether LLM agents internally represent the dependency graph of their tool calls. Using low-capacity edge probes on the model’s residual streams, the authors find that tool-call dependency structure is linearly decodable – meaning the model maintains an internal representation of which earlier tool outputs feed into later calls. This has practical implications for agent monitoring: if the model’s internal dependency graph diverges from the intended execution plan, that divergence may be detectable from activations alone.

Reasoning efficiency: how much thinking is actually necessary? #

arXiv (2605.23926, 2605.25745)

Two papers address the cost of reasoning from complementary directions. “How Much Thinking is Enough?” measures at scale the redundancy in LLM reasoning chains – the extensive reformulation, verification, and circular self-reflection that inflates inference cost without improving accuracy. “Selective Latent Thinking” proposes adaptive compression that preserves precision-critical reasoning steps while compressing easy ones into latent representations. Together they sketch a practical path toward cheaper reasoning inference without sacrificing accuracy on hard problems.

ContextEcho: Persona Drift in Long Agentic-Coding Sessions #

arXiv

A model’s “helpful programming assistant” persona does not survive long agentic-coding sessions. After hours of tool-using debugging, models that initially hedge preferences begin asserting them, shift communication styles, and exhibit measurable persona drift in the deployment regime that production coding products actually use. For teams running persistent coding agent sessions, this identifies a reliability concern that session-length limits or periodic context resets may need to address.

Security #

When the Manual Lies: MCP Poisoning Attacks for LLM Agents #

arXiv

The Model Context Protocol (MCP) standardizes how LLM agents discover and invoke external tools, but this paper demonstrates that the protocol’s tool descriptions can be poisoned to hijack agent behavior. By modifying the “manual” (tool descriptions) that agents rely on, attackers can redirect tool calls, exfiltrate data, or trigger unintended actions without modifying the agent’s code or model weights. The attack surface is concerning because MCP tool registries are often populated from external sources that the agent operator does not fully control.

Inference-Time Backdoors via Chat Templates: From LLM Supply Chains to Agentic System Compromise #

arXiv

Open-weight models can carry backdoors embedded not in model weights but in chat templates – the formatting layer that converts conversations into model input. Since chat templates are typically trusted and rarely audited, this creates a supply chain attack vector: a compromised template on a model hub activates hidden behaviors only when specific trigger conditions are met during inference. For agentic systems that assemble multiple open-weight models, each model’s chat template is an unaudited trust boundary.

Regulatory & Policy #

A reality check on the AI jobs hysteria #

MIT Technology Review

MIT Technology Review pushes back against the narrative that AI is decimating white-collar employment, arguing that while individual company layoffs generate headlines, aggregate employment in developed countries remains broadly stable. A companion piece on the crisis in entry-level work argues the real threat is not mass unemployment but the erosion of entry-level positions that serve as training grounds – companies are automating junior tasks faster than they are creating new pathways for workers to develop the judgment that AI cannot yet replicate. The structural risk is a workforce that loses its pipeline for developing expertise.

Infrastructure #

Norway’s 2 petabytes of Huawei flash storage and LLM training #

Blocks & Files / Hacker News

Norway is deploying 2 petabytes of Huawei flash storage infrastructure for LLM training workloads. While the US and allies have restricted Huawei chip exports, storage infrastructure occupies a different regulatory position, and Norway’s choice highlights how AI training dependencies extend beyond GPUs to the broader storage and data pipeline stack. For teams planning AI infrastructure, storage bandwidth and capacity are increasingly constraining factors alongside compute.

An Agent-Driven End-to-End HW-SW Co-Design Benchmark for Heterogeneous SoCs #

Semiconductor Engineering (Columbia University, IBM Research)

Researchers from Columbia and IBM released HSCO-Bench, a benchmark for evaluating AI agents that perform end-to-end hardware-software co-design for heterogeneous systems-on-chip. The benchmark tests whether agents can navigate the full design space from architecture specification through synthesis and simulation, a task requiring reasoning across abstraction layers that traditional EDA tools handle in isolation. This represents agentic AI expanding from software-only tasks into hardware design workflows where the cost of errors is measured in silicon respins.

Open Source #

The Open/Closed Problem in AI #

Lobsters

This essay argues that the AI industry is hardening around open-loop learning (train offline, deploy frozen weights) because inference ASICs physically embed this assumption – weights are read-only at the silicon level. Drawing a parallel to GPU history (fixed-pipeline to programmable and back), the author warns that the window to experiment with closed-loop learning (models that update from prediction errors, analogous to biological brains) is closing with every specialized chip that ships. For infrastructure planners, this frames ASIC specialization not just as an efficiency win but as a paradigm lock-in with long-term research consequences.

Threads to Watch #

The harness thesis is crystallizing. A Hugging Face terminology guide, two independent arXiv papers, and Nolan Lawson’s coding workflow all converge on the same insight: the infrastructure layer around a foundation model – the harness – determines more of the system’s behavior than the model itself. “Stop Comparing LLM Agents Without Disclosing the Harness” calls for mandatory disclosure in benchmarks. “From Model Scaling to System Scaling” argues harness architecture should be a first-class engineering discipline. This follows last week’s pattern of agent failure-mode research and suggests the field is decisively shifting investment from model capabilities to system-level design.

AI labor substitution narratives meet macro data. ClickUp’s 3,000-agent deployment alongside a 22% workforce reduction is the most explicit corporate case yet of framing layoffs as AI substitution. MIT Technology Review’s same-day analysis argues aggregate employment remains stable and the real risk is the erosion of entry-level pipelines. The tension between company-level restructuring stories and economy-wide stability data will define how organizations and policymakers respond to AI workforce transformation through 2026.

Sources Unavailable Today #

These sources could not be fetched today. Links point to their homepages so you can check them directly.

• Stanford HAI — scrape: content_truncated