Why does rethinking the workflow around an AI agent pay off more than making the agent itself smarter?
How does workflow redesign create more value than improving agent sophistication?
This explores why companies get more out of AI agents by rethinking how the work itself is organized than by building smarter, more capable agents, and what research beyond the consulting reports says about where agent value actually comes from.
This explores why redesigning the work tends to beat upgrading the agent. The consulting evidence is blunt. McKinsey looked at more than 50 agent builds and found that teams who mapped the whole workflow, found the real pain points and fitted agents to specific steps got results. Teams that focused on making the agent impressive often ended up with tools that looked capable but didn't change outcomes Does agent quality matter more than workflow redesign?. Deloitte describes the same failure from the other side: agents added onto old processes stall. Organizations that succeed rebuild the process end to end, with data agents can use, orchestration and new ways of managing the work Does agentic AI succeed through process redesign or workflow layering?.
The surprising part is that research on how agents are built reaches the same conclusion from below. Reliable agents don't come mainly from bigger models. They come from moving memory, reusable skills and interaction rules out of the model and into the system around it, the 'harness', so the model doesn't have to solve the same problems again on every run Where does agent reliability actually come from?. Skills seem to help mostly by giving the agent a stable procedure to follow, not by supplying missing facts. Procedural anchoring explains about two-thirds of the cases where skills helped, while supplying knowledge explains under 5% Do skills teach procedures or inject missing facts?. In other words, a well-designed workflow written into the system does much of the work people expect intelligence to do. Agents can even learn this structure themselves: pulling reusable sub-task routines out of past runs raised success rates by 24–51% Can agents learn reusable sub-task routines from past experience?.
Redesign also means not using an agent everywhere. McKinsey's guidance is to save agents for messy, unpredictable work and use simpler automation for routine, standardized steps. Defaulting to agents adds complexity without matching value When should you actually deploy an AI agent instead of simpler automation?. Builders report the same thing. Teams running agents in production swapped flexible tool protocols for explicit function calls, one tool per agent, because predictable behavior mattered more than flexibility Why do protocol-based tool integrations fail in production workflows?. Cost follows the same pattern: most agent sub-tasks are repetitive enough for small models at 10–30× lower cost Can small language models handle most agent tasks?, and a mid-sized model trained on real multi-step execution can compete with much larger ones on cost and performance Does model efficiency matter more than peak capability for real work?.
Here's the twist you might not expect: past a certain point, a more capable model can get less out of a good workflow. When agents were given improved harnesses, the benefit peaked in mid-tier models. Weak models didn't use the harness, and the strongest ones were worse at following its instructions faithfully Do stronger models always evolve harnesses better?. Workflow and capability aren't simple substitutes, and adding more of one doesn't automatically add to the other.
An open question is who redesigns the workflow. One approach keeps people involved: letting users assign plan steps to themselves or the agent, and revise the plan mid-task, gave them more control without extra effort Can users steer AI agents better by revising plans mid-execution?. Another hands the job to AI: meta-agents trained with reinforcement learning build a custom multi-agent workflow for each query Can AI systems design unique multi-agent workflows per individual query?. One caveat: the strongest 'workflow beats sophistication' claims come from consulting reports, not controlled studies. The research on harnesses and skills supports them from another angle, but nobody here measures the trade-off directly.
Sources 12 notes
McKinsey's analysis of 50+ agentic AI builds shows that organizations gain value by redesigning entire workflows—mapping steps, identifying pain points, and fitting agents to specific tasks—rather than maximizing agent sophistication alone. Organizations that focus too heavily on the agent itself often end up with capable-looking tools that fail to improve overall results.
Deloitte argues many agentic AI implementations fail because organizations layer agents onto old workflows rather than redesigning operations. Success requires agent-compatible data architectures, robust orchestration, and new management models; HPE and Toyota illustrate this approach.
Research shows reliable LLM agents externalize three cognitive burdens—memory (state persistence), skills (procedural components), and protocols (structured interaction)—into a harness layer rather than relying on model scale alone. The harness unifies these externalities and eliminates the need for the model to solve the same problems repeatedly.
Analysis of 8,135 trials shows procedural anchoring accounts for 65.7% of skill cases versus 4.5% for knowledge injection. Skills fail when retrieved incorrectly, invoked out of context, or followed too rigidly.
Agent Workflow Memory induces sub-task routines at finer granularity than full tasks, abstracts example-specific values, and compounds them hierarchically. This produces 24.6% relative gain on Mind2Web and 51.1% on WebArena, with larger gains as train-test gaps widen.
Show all 12 sources
McKinsey's framework recommends reserving agents for high-variance, low-standardization work like financial document extraction, while using simpler automation for predictable, standardized tasks. Defaulting to agents adds complexity without proportional value in well-governed workflows.
MCP integration caused non-deterministic failures through ambiguous tool selection and parameter inference. Replacing it with explicit direct function calls and single-tool-per-agent design restored determinism. A 306-practitioner survey confirms 85% of production teams build custom agents, forgoing frameworks.
SLMs handle the repetitive, well-defined language tasks that constitute most agent work at 10–30× lower cost than LLMs, making heterogeneous architectures (SLMs by default, LLMs selective) the economically rational design pattern.
Occamy-1.0, a 35B-parameter model further trained on execution-grounded data and long-horizon trajectories, achieves competitive performance with much larger models while sitting at the low-cost knee of the Pareto frontier, suggesting that training for coordination and follow-through substitutes for raw scale in multi-step work.
Model capability to produce useful harness edits stays constant across tiers, but capacity to actually benefit from those edits follows an inverted U-shape, peaking in mid-tier models. Weak models fail to invoke harnesses; strong models struggle with faithful instruction-following.
Cocoa's system, which lets users assign plan steps to themselves or an agent and revise plans mid-execution, afforded greater steerability than chat-baseline interaction in a lab study (n=16) while maintaining ease of use, with field users valuing explicit work delegation.
FlowReasoner demonstrates that meta-agents trained with reinforcement learning and external execution feedback can generate unique multi-agent architectures for each user query, optimizing across performance, complexity, and efficiency—moving beyond fixed task-level workflow templates.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Harness Updating Is Not Harness Benefit: Disentangling Evolution Capabilities in Self-Evolving LLM Agents
- One year of agentic AI: Six lessons from the people doing the work
- Why Do Multi-agent LLM Systems Fail?
- Unraveling Human-AI Teaming: A Review and Outlook
- Demystifying Agent Skills: Why They Work-Until They Don't
- Towards a Science of Scaling Agent Systems
- A Practical Guide for Designing, Developing, and Deploying Production-Grade Agentic AI Workflows
- RRSI: Regularized Recursive Self-Improvement of Agent Harnesses