Is the hard part of office work brainpower, or the messy coordination around it, like people, software, and know-how?
What fraction of real workplace tasks require frontier-scale reasoning versus coordination?
This explores whether the hard part of real workplace tasks is raw reasoning power or the organizing work around it (coordination, social skill, domain knowledge). The corpus has no measured fraction, but it shows where agents actually break.
This explores whether the hard part of real workplace tasks is raw reasoning power or the organizing work around it. No note in the collection sorts tasks into 'needs frontier reasoning' and 'needs coordination', so there is no percentage to give. The indirect evidence does point one way, though. On TheAgentCompany, a simulated office, leading agents finish only about 30% of tasks, and the three named failure modes are social interaction, navigating professional software, and domain-specific knowledge Why do AI agents fail at workplace social interaction?. None of those looks like a math-olympiad problem.
Frontier-scale reasoning is also less tied to size than it sounds. A 3B model trained with the right post-training pipeline matches much larger systems on hard math and coding benchmarks. That holds only for tasks with checkable answers, where reinforcement learning gets a clean reward signal Can small models match frontier reasoning without massive scale?. Much office work has no answer key, so reasoning is cheapest exactly where workplaces are least like a benchmark. Chain-of-thought also degrades predictably when tasks drift from the training distribution, producing fluent reasoning that isn't logically sound Does chain-of-thought reasoning actually generalize beyond training data?. Extra reasoning training can even hurt: it helps math but can degrade medicine, because knowledge and reasoning live in different parts of the network Why does reasoning training help math but hurt medical tasks?.
On the coordination side, the argument is that some limits can't be fixed with a smarter model. Tasks that need different kinds of expertise, parallel work, and independent checking exceed what any single agent loop can organize, however capable it is Do single agents always hit organizational limits?. Splitting the work up turns out to be a more portable skill than doing it. A separate decomposer model transfers across domains, while the solver model does not Does separating planning from execution improve reasoning accuracy?. Coordination itself isn't solved. Agents that pass standardized documents, as in MetaGPT, coordinate better than agents that chat Does structured artifact sharing outperform conversational coordination?. Larger agent networks still fail by agreeing too late or by accepting neighbors' claims without checking them Why do multi-agent systems fail to coordinate at scale?.
The split may also be blurrier than the question assumes. Some of the best reasoning gains come from importing coordination structure into a single model. Spreading test-time compute across diverse abstractions beats sampling more solutions in parallel Can abstractions guide exploration better than depth alone?, and a model that reasons as a dialogue between distinct agents beats monologue reasoning Can dialogue format help models reason more diversely?. So most workplace tasks probably need a little of both. Reasoning gains are concentrated in the checkable slice, and the failures the corpus documents sit mostly in people, tools, and organization. Getting a real fraction would take a benchmark that labels each failed task by its cause, and the collection doesn't have one yet.
Sources 10 notes
TheAgentCompany benchmark shows leading agents achieve 30% task completion in a simulated workplace. Social interaction, professional UI navigation, and domain-specific knowledge are the three primary failure modes, with multi-turn task performance consistently dropping to 35% across enterprise settings.
A 3B model trained with curriculum SFT and multi-domain RL reaches 94.3 AIME26 and 80.2 LiveCodeBench scores matching much larger systems. The result is bounded to verifiable tasks with checkable ground truth, where RL can provide clean reward signals.
DataAlchemy experiments show CoT fails systematically under distributional shifts in task, length, and format. Models produce fluent but logically inconsistent reasoning — imitating reasoning form without valid underlying logic.
Two-phase inference model shows knowledge retrieval operates in lower network layers while reasoning adjustment happens in higher layers. This separation explains why reasoning training improves math but can degrade knowledge-intensive domains like medicine.
Research shows that real-world tasks requiring heterogeneous expertise, parallel execution, and independent verification exceed what any single agent loop can organize. Graph-based system abstractions are needed to distribute intelligence across specialized agents.
Show all 10 sources
Modular architectures with separate decomposer and solver models outperform monolithic LLMs, with decomposition ability transferring across domains while solving ability does not. The separation prevents planning-execution interference and produces more generalizable skills.
MetaGPT demonstrates that agents producing standardized engineering documents achieve superior coordination compared to conversational exchange. Active information pulling from shared environments eliminates noise and mirrors efficient human workplace infrastructure.
AgentsNet benchmark shows agents fail to coordinate strategies either by agreeing too late or adopting strategies without informing neighbors. Agents accept neighbor information without verification, enabling error propagation while remaining capable of detecting direct conflicts.
RLAD jointly trains abstraction and solution generators, showing that allocating test-time compute to diverse abstractions outperforms parallel solution sampling at large budgets. Abstractions create structured breadth-first exploration that prevents the underthinking failure mode of depth-only reasoning chains.
DialogueReason, which structures a single model's internal reasoning as dialogue between distinct agents in separate scenes, overcomes monologue reasoning's fixed-strategy and fragmented-attention weaknesses, especially on tasks requiring multiple problem-solving approaches.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Towards a Science of Scaling Agent Systems
- Drop the Hierarchy and Roles: How Self-Organizing LLM Agents Outperform Designed Structures
- Is Chain-of-Thought Reasoning of LLMs a Mirage? A Data Distribution Lens
- DialogueReason: Rule-Based RL Sparks Dialogue Reasoning in LLMs
- Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?
- Reasoning LLMs are Wandering Solution Explorers
- A Comment On "The Illusion of Thinking": Reframing the Reasoning Cliff as an Agentic Gap
- From Model Scaling to System Scaling: Scaling the Harness in Agentic AI