Successful agent adoption comes from teaching engineers how to delegate, review, and gradually internalize practices, rather than relying on one fixed framework.
2
Coding agents are becoming a general software primitive, while faster inference, specialized models, agentic memory, and agent-ready product interfaces change how software is built and used.
3
MCP and command-line tools each work well in different situations, so production agents will often need both rather than a single universal integration pattern.
Summary
Day 2 of AIE Miami covered the engineering work around coding agents, inference, memory, interfaces, and adoption. David House described G2i case studies where staged documentation, reviews, testing, and mentorship helped engineers move from hesitation to confident delegation. Sarah Chieng argued that growing context and output sizes created "latency debt," and showed how faster inference changes coding agents from tools developers leave running into pair programmers. Other talks covered mobile diffusion models, specialized sub-agents, coding agents as general software infrastructure, graph-based context, and memory systems that let agents reason over raw session data. The afternoon shifted toward products and users. Speakers discussed building AI apps with simple architectures, open-source model choice, agent-ready APIs and pricing, and the runtime required for robots and multimodal systems. Laurie Voss presented an evaluation comparing MCP with GitHub's CLI. David Gomes described an IDE that keeps a code editor while adding richer interfaces for supervising cloud agents.
Agent adoption improves when workflows teach engineers how to delegate
David House's G2i case studies followed engineers who began with different levels of trust in AI. Ava learned through a project where agent use was expected and eventually wrote a specialized testing sub-agent. Lucy first produced duplicated, hard-to-review code, then learned to require tests and test runs. Dale created a 10,000-line pull request before learning to narrow the scope of delegation. House said an agent framework should constrain beginner input and amplify expert input. At G2i, the framework used product briefs, technical specifications, code or test-driven development, review, and pull request skills. Its purpose was to reveal hidden practices, make work reviewable, and teach delegation.
Faster inference changes coding agents from background jobs into pair programmers
Sarah Chieng called the hidden cost of larger models, longer prompts, and longer outputs "latency debt." She said model speed had generally stayed between 50 and 150 tokens per second while input lengths grew four times and output tokens grew three times in the cited OpenRouter and A16Z study. Chieng presented Codex Spark, which she said can generate 1,200 tokens per second, and argued that this makes it practical to sit with the model, verify its output, and steer it. She connected the speed gains to hardware, on-chip memory, disaggregated prefill and decode, mixture-of-experts models, pruning, and KV-cache reuse.
Mobile NPUs can run small diffusion experiences without cloud calls
Lech Kalinowski described deploying a latent diffusion model directly on a phone's neural processing unit. His system removed the usual text-to-embedding prompt path and used an ambient sensor value, such as light, to condition the latent update. The application used a native Android bridge and ONNX Runtime, with an encoder, a denoising loop of 8, 16, or 24 steps, and a decoder that produced an RGB image. He demonstrated the phone in flight mode and reported roughly 600 milliseconds of inference latency. A thermal fallback could stop the application if the device ran out of memory or became too hot. He suggested using NPUs for unconventional local applications and animations.
Specialized sub-agents reserve frontier models for work that needs them
Tejas Bhakta described "Software 3.5" as agents prompting other agents, with each sub-agent using a separate context window. Morph uses specialized models for tasks such as code search, context compaction, and applying edits, while keeping coding, planning, and reasoning on frontier models. Bhakta said Morph's code-search model can make up to 12 parallel reads or directory listings over six turns, and its compaction model runs at 33,000 tokens per second. He discussed speculative decoding, disaggregated prefill, and custom kernels as inference techniques. For sub-agents, he recommended small outputs, shared prefix caches for task agents, and training models to behave predictably inside a larger agent system.
A coding agent can become the general software primitive for business work
Rick Blalock argued that coding agents now do more than write code for developers. They can build chatbots, use retrieval, orchestrate other agents, create databases, and write tools for themselves when given the right instructions. He contrasted early agent frameworks, which often looked like deterministic chains, with newer systems that give agents a body through sandboxes, persistence, integrations, and observability. Blalock described businesses using coding agents for marketing, sales estimates, and operations, including a window-cleaning company and a construction company. His conclusion was that teams choosing an agent framework should consider starting with a coding agent that has an SDK, server-side handling, resumable sandboxes, and observability.
Graphs add the relationships and decision traces that text retrieval misses
Nyah Macklin used a banking example where separate retrieved facts about Jessica, her employer, sanctions, and a credit request did not connect in vector search. She argued that vector search captures textual meaning while graph structure captures how entities are related. Her example represented an apple as a human concept, an embedding vector, and a knowledge graph with explicit nodes and relationships. Macklin cited a 2026 telecom study where a base model reached 37% accuracy, fine-tuning reached 54%, and combining a knowledge graph with retrieval-augmented generation reached 91%. She then described context graphs as knowledge graphs that record full decision traces, causal relationships, policies, and the context behind significant agent decisions.
Agent products need a behavior runtime between model intent and real-world action
Lena Hall's Reachy Mini robot exposed the coordination problems in multimodal systems. Her architecture separated the physical robot, local media pipelines, real-time orchestration, tools and motion, and profile and personality configuration. Hall insisted that the model should express intent through a tool call while the runtime decides whether and how to move hardware. She also used a response worker to serialize speech and aggressively clear audio and motion state after interruptions. In her examples, audio came online in 400 milliseconds, the real-time session became ready in 2.3 seconds, and a camera result took more than four seconds to appear in the final response. She argued that timing, interruption handling, tool boundaries, personality, and observability are product behavior.
Agents are first-class product users with different access and pricing needs
Dave Kiss argued that agents increasingly become the users of SaaS products, so products need interfaces designed for them. He discussed agent access through APIs, MCP, CLIs, structured error responses, pricing data, and machine-readable documentation. He cited GitHub's projected growth from about one billion commits in 2025 to 14 billion in 2026 if the trend continued, and said agents have less patience for product friction than humans. Kiss suggested publishing a discoverable pricing Markdown file, giving errors actionable next steps, generating a CLI from an OpenAPI specification, and testing your own product through an agent. He also described exposing video data such as transcripts, summaries, and metadata so agents can work with it.
Memory systems must manage state, curation, and change over time
Alvin Pane described dynamic memory discovery, an approach that used raw JSON session files, an agent orchestrator, file-system tools, recursive calls to other agents or language models, and a plan file. He said this simple design reached 100% on LongMemEval after earlier results had already been surpassed. Pane was honest that a benchmark does not solve production memory. He separated memory into state, which gives context somewhere to live; curation, which selects what enters the context window; and life cycle, which handles updates, changing facts, and forgetting. He also named temporal reasoning, entity disambiguation, and principled forgetting as unresolved problems. His view was that every agent system is already building memory, whether its designers call it that or not.
"An agentic framework should shape the input of the human in addition to improving the agent's output."08:39
Who should watch
You are introducing coding agents to engineers who need a practical adoption path, review process, or mentorship model.
Your agent system is slow, forgetful, hard to audit, or connected to tools through interfaces designed only for humans.
You are choosing between MCP, a CLI, specialized models, graph retrieval, or a new agent-oriented runtime and want evidence from the talks rather than slogans.
