The model is only one part of the agent.
Dialogo improved every model evaluated to date by 8-12 percentage points on autonomous tool-use tasks, without fine-tuning or changing model weights.
The result
A stronger harness changed the outcome. The model stayed the same.
On MCP-Atlas Public-500, Dialogo with Gemini 3 Flash reached approximately 74.4%, compared with the published Gemini 3 Flash baseline of 64.0%.
That is a 10.4 percentage-point uplift, or 16.25% relative improvement.
Figure 01
Pass-rate uplift with the same executor
Published Gemini 3 Flash baseline versus Dialogo with the same executor. Dialogo result is self-run and not an official Scale leaderboard entry.
Figure 02
Successful tasks completed
At the displayed pass rates, the 500-task public set translates to 320 baseline passes and 372 Dialogo passes.
What the uplift means
52 more successful outcomes across 500 tasks.
A pass rate can feel abstract. On the public 500-task set, 64.0% corresponds to 320 passed tasks and 74.4% corresponds to 372.
This task-count view assumes the displayed rates represent exactly 320 and 372 passes. The raw scorer output should remain the publication source of truth.
The wider field
A smaller model can compete in a bigger-model performance band.
With Dialogo, Gemini 3 Flash moves from its 64.0% published baseline to 74.4%. In this evaluation, that puts the efficient model in the same performance band as several substantially larger frontier models.
Public-500 comparison
Dialogo in the full published field
A focused comparison
Flash economics, frontier-range tool use.
The selected view makes the immediate performance neighborhood easier to read. The complete chart above remains the authoritative on-page comparison across every published Public-500 score.
Figure 03
Selected Public-500 context
Selected official Scale Labs Public-500 values. Dialogo is independently measured and separately labeled for positioning context.
What Dialogo is
A model-independent execution layer
A frontier, efficient, specialized, or self-hosted open-weight model can remain the executor. The model can be replaced without rebuilding the complete workflow.
An agent harness around the model
Dialogo manages planning, tool use, context, recovery, state, verification, permissions, and human approvals without changing model weights.
Where the uplift comes from
Intelligence is a property of the whole execution system.
Plan and track subgoals
Break a goal into accountable steps, preserve execution state, and detect incomplete requirements.
Use tools with control
Discover tools, generate schema-aware arguments, manage outputs, and recover from transient failures.
Verify before delivery
Check intermediate results, validate the final answer, and pause consequential actions for approval.
Optimize successful work
Route by complexity, cap unnecessary calls, and measure cost against completed tasks instead of tokens alone.
The model produces decisions. The harness controls how those decisions become reliable actions.
The efficiency question
Measure cost per successful task, not cost per token alone.
A smaller model is only economical if it completes the work. A stronger harness can route simple steps efficiently, reserve expensive reasoning for difficult decisions, limit oversized outputs, and recover without restarting the complete task.
The chart is a normalized mathematical scenario, not measured billing. Exact savings require matched token, tool, and runtime logs.
Figure 04
Normalized cost-per-success break-even
Mathematical break-even analysis based on 320 versus 372 passed tasks. This is not a measured billing comparison.
Research context
Agent architecture transfers across models.
Sakana AI's Fugu explores learned model orchestration through selection, delegation, verification, and synthesis. Darwin Gödel Machine and ShinkaEvolve provide further evidence that improvements to an agent scaffold can generalize across underlying models.
Learned orchestration for model selection, delegation, verification, and synthesis.
Darwin Godel MachineReported 20.0% to 50.0% on SWE-bench and 14.2% to 30.7% on Polyglot while improving the agent system.
ShinkaEvolveAn evolved agent scaffold evaluated for generalization to unseen tasks and models.
System-level intelligence depends not only on the model, but on how models, tools, context, verification, and execution are orchestrated.
Methodology and evidence
What the current evaluation shows.
MCP-Atlas evaluates realistic, multi-step tool use. The public split contains 500 tasks across real MCP servers, and tasks pass when claim coverage reaches the benchmark threshold.
Measured results
- Every model evaluated by Dialogo to date improved.
- Observed gains range from 8 to 12 percentage points.
- Gemini 3 Flash showed a 10.4-point uplift against the cited public baseline.
- No underlying model weights were modified.
Operational implication
The measured Gemini 3 Flash configuration reached a performance band occupied by larger frontier models. This supports a practical strategy: improve the orchestration system first, then select the smallest executor that meets the workflow's quality, latency, privacy, and cost requirements.
A model call is not an autonomous system.
The model matters. So do the plan, tools, context, recovery, permissions, and verification that turn a response into completed work.