aime™ extends its Smarter-on-Smaller-Models programme — early results across the agent stack

On May 29, aime set out the engineering posture behind its Smarter-on-Smaller-Models programme: only ask a model to do what only a model can do, and move counting, lookup, sequencing, validation and arithmetic into deterministic code. Today aime is reporting continued progress against that programme as it moves through the agent stack.

Where the programme stands

The first wave of agents has now been re-shaped around small-model-friendly patterns. Curriculum extraction, planning and recommendation have each been decomposed into narrower responsibilities, with deterministic software taking over the work that does not require language reasoning. The model continues to do what it is uniquely good at — generating human-quality instructional content — while structure, identifiers, sequencing and computation move into code.

• Curriculum extraction is now a flat, single-responsibility pass — hierarchy is reconstructed deterministically downstream.
• Planning has been split into sequenced passes, each with smaller prompts and tighter schemas.
• Recommendation logic has been separated cleanly from instructional phrasing — rules in code, language in the model.
• Validation, repair and structured output are standardised through aime's Kern™ orchestration layer.

Continuing innovation

The next phase focuses on the Lesson Designer — historically the most complex agent in the platform — and on the assessment and feedback paths. The direction is the same throughout: smaller responsibilities, smaller prompts, smaller schemas, and a clearer division of labour between the model and the surrounding software. The specific prompt structures, schema decompositions, repair strategies and small-model mitigations developed in the course of this work are aime's proprietary IP and are shared with partners under NDA.

“The most interesting result so far is not the cost curve — it is the reliability curve. When a small model is given a smaller job, it does that job better, more consistently and more cheaply than a larger model asked to do everything at once.”

Why it matters

The platform-level target remains a 60–80% reduction in LLM cost and prompt complexity across the agent stack, with classroom-readiness held constant against aime's existing validators. The strategic outcome is the same: a system that can run a different lesson pack for a teacher in rural India and a different one again for a teacher in the UK — both for pennies, both classroom-ready, both deployable on modest infrastructure.

“Continuing innovation here looks unglamorous from the outside. It is not a bigger model. It is a quieter one, doing less, more reliably, inside a better-designed system. That is what makes educational intelligence affordable at population scale.”

Availability

The optimised agents are rolling out across aimeCLOUD™ as each pass completes internal validation. The remaining phases — Lesson Designer micro-agents, assessment decomposition and the feedback path — are sequenced inside the aime Loom™ Workflow Engine for deterministic, replayable execution.

About aime

aime builds the operating system for educational intelligence — the foundational infrastructure layer that future education systems will run on. aime's stack combines structured curriculum knowledge, pedagogy-aware reasoning, compact education-tuned models, agentic orchestration and offline-capable deployment, and is designed for ministries of education, universities and national school systems.