Effort Scaling and Agent Orchestration: A New Lever for Multi-Agent Scale
Around September 30, a concentrated wave of discussion formed around effort scaling—having models allocate reasoning intensity according to task difficulty—widely viewed as a key new lever for scaling multi-agent systems. Elvis Saravia (omarsar0) published a dense series of diagrams on agent architecture and orchestration, covering how agents scale, staying on track, manager patterns, and agent team patterns; related threads covered loop engineering (telling agents when to stop) and the view that harnesses shape model optimization in reverse.
Confirmed
- Elvis Saravia (omarsar0) repeatedly argued that the biggest current challenge is scaling multi-agent systems, and that effort scaling helps orchestrate subagents better. His basis was Anthropic's engineering blog on two orchestration mechanisms: effort scaling and a companion agent coordinating skill; he also relayed the advice "Don't over manage your agents" and recommended testing that coordinating skill.
- Elvis published a diagram series: first on keeping agents on track with a manager pattern, then a follow-up on how agents scale; the agent team pattern aims to improve inter-agent communication and support deeper exploration, with more detail promised in coming days based on his own research.
- Lawrence Chan's team's /goal scaling experiments observed that pushing the agent harness to do more improved the model's long-horizon capabilities; a core difficulty is telling agents when to stop—the loop engineering that Seymour discussed in depth.
- jzhao's team introduced two Agent updates: dynamically choosing its own reasoning effort (less compute on easy tasks, more on hard ones) and improved handoffs when delegating to subagents. The authors claim it tops the cost-efficiency frontier on DeepSWE and Terminal-Bench 4.0.
- Elvis live-noted a tech talk opening with the Codex agent and its harness, stressing that the harness in turn determines how the model gets optimized—training objectives are inseparable from the agent's runtime environment.
Why it matters
- The shared thrust: rather than simply adding more compute or more agents, systems should learn to allocate reasoning effort on demand and gain stronger orchestration and stopping judgment—potentially a more economical path to better long-horizon performance.
- The two-way relationship between harness and model optimization means agent engineering and model training are deeply coupling, worth practitioners' attention.
- Continued primers and experiments on agent team and manager patterns show multi-agent communication and exploration mechanisms are an active practical topic.
2026-09-30 ~ 2026-09-30 · 12 related posts
Primary sources
- Agent Now Picks Its Own Effort and Hands Off to Subagents, Leading Cost-Performance on DeepSWE and Terminal-Bench 4.0 — _jzhao ·
- Effort scaling seen as key lever for orchestrating multi-agent systems — omarsar0 ·
- Effort scaling plus a coordinating skill promises better subagent orchestration — omarsar0 ·
- [source] Agent Now Picks Its Own Effort and Hands Off to Subagents, Leading Cost-Performance on DeepSWE and Terminal-Bench 4.0 — _jzhao · 2026-09-30
- Key notes from a talk on the Codex agent and how harnesses shape model optimization — omarsar0 · 2026-09-30
- From 'when to stop' to loop engineering: agent harnesses shape model training — omarsar0 · 2026-09-30
- Loop engineering: pushing the agent harness harder boosts long-horizon model capabilities — omarsar0 · 2026-09-30
- Scaling multi-agent systems is the next big challenge, harness gains lift model capability — omarsar0 · 2026-09-30
- [source] Effort scaling seen as key lever for orchestrating multi-agent systems — omarsar0 · 2026-09-30
- [source] Effort scaling plus a coordinating skill promises better subagent orchestration — omarsar0 · 2026-09-30
- Don't over manage your agents: a coordinating skill worth testing — omarsar0 · 2026-09-30
- Elvis Saravia Illustrates How to Keep AI Agents on Track with Manager Pattern — omarsar0 · 2026-09-30
- Elvis Saravia Shares New Diagram on How to Scale AI Agents — omarsar0 · 2026-09-30
- Agent team pattern improves inter-agent communication and deep exploration — omarsar0 · 2026-09-30
- Researcher: agent team pattern aims to improve inter-agent communication and deep exploration — omarsar0 · 2026-09-30