An overview of the open-source AI development stack for building agentic coding workflows, broken into the 'MIGHT' layers: Model, Inference, Gateways/routers, Harness, and Tools (skills/MCP). Covers when to use large vs. small open models (e.g., Kimi K3 vs. GLM 5.3 Flash), inference providers like Together AI, gateways like OpenRouter and Vercel AI Gateway, harnesses like OpenCode, PI, and Amp, and practices like managing context, starting fresh sessions, and a plan-implement-review workflow across multiple models. Argues the key advantage of open models is a composable stack where each layer can be swapped independently.
Table of contents
The MIGHT StackModelsInference providersGateways and routersHarnessesTools (skills and MCP)Build for experimentationThe open stackQuestions this post answers
When should I use a large model like Kimi K3 versus a small model like GLM 5.3 Flash for coding tasks?
Use large models such as Kimi K3 for ambiguous, multi-step work like refactoring authentication systems, upgrading frameworks, or reviewing pull requests, since their extra capacity handles unclear requirements well. Use small models such as GLM 5.3 Flash for narrowly scoped tasks like adding a function option or writing tests, since they are roughly 6 times smaller and 20 times cheaper than Kimi K3 while matching its performance on well-specified work. daily.dev helps developers compare model tradeoffs like this when picking tools for a coding stack.
What is the difference between an AI gateway and a harness in an AI coding agent stack?
A gateway sits in front of multiple inference providers, aggregating them behind one API so requests can be routed, compared on price and latency, and switched without code changes; examples include OpenRouter, Vercel AI Gateway, and self-hosted LiteLLM. A harness is the application layer that manages the conversation, gives the model tools to search code, read files, and apply patches, and connects it to a codebase; examples include OpenCode, PI, and Amp. developers assembling their own agent stack track these layer distinctions on daily.dev.
What is a plan-implement-review workflow using multiple AI models for coding?
It splits a coding task across three model calls: a large model plans by breaking an open-ended prompt into isolated tasks, a small model implements each task in its own fresh session to keep context focused, and a large model reviews the completed work, looping back to planning if problems appear. This lets teams use cheaper small models for routine implementation while reserving larger models for reasoning-heavy planning and review. daily.dev keeps developers experimenting with multi-model agent workflows like this one moving.
Share this post