Topic
Agents and multimodal systems
The infrastructure under agent loops, retrieval pipelines and image, video and audio workloads, where one request becomes many.
14 articles · Last updated
Articles
Agents & multimodal
AI image and video production for agencies
Generative tools multiply inside an agency faster than anyone can track them. This guide covers whose provider contract a job runs on, how to price a job before it runs, how to compare takes across models, how review should work, and what to do when a vendor retires the model a campaign was built on.
Guide11 min read
Agents & multimodal
INFRO Studio: image, video and audio on your own providers
Media models bill per image, per second or per character, run as async jobs, and often delete their outputs within hours or days. Here is how those APIs differ, and how INFRO Studio puts projects, takes, estimates and storage on top of the provider accounts you already pay.
Guide12 min read
Agents & multimodal
MCP gateway: one governed entry point for agent tools
Every MCP server an agent can reach is code and credentials you now depend on. This guide covers the incidents that made that clear, what a gateway in front of your tool servers can and cannot control, and the INFRO MCP gateway's endpoint, security model, limits and errors.
Guide12 min read
Agents & multimodal
Image models compared: GPT Image 2, Nano Banana, FLUX.2
Image models are billed three different ways — per token, per image and per megapixel — and two well-known ones were retired this year. A comparison on documented facts, and a way to decide on your own prompts.
Comparison7 min read
Agents & multimodal
AI video models compared: Veo 3.1, Kling, Wan and Sora's exit
Sora 2 and OpenAI's Videos API shut down on 24 September 2026. What the remaining video APIs document about clip length, resolution, audio, asynchronous jobs, output retention and per-second rates — and how to choose without a leaderboard.
Comparison8 min read
Agents & multimodal
Async media jobs: webhooks, idempotency and retries
Video renders take minutes and bill per second, so a duplicate submit or a missed completion costs real money. This guide covers idempotent submits, verifying signed webhooks, deduplicating at-least-once deliveries and recovering from lost events.
Guide9 min read
Agents & multimodal
Bounding AI agent loops: step, time, token and spend limits
An agent that loops is a bill that grows quadratically. This guide builds a run guard for step, time, token and spend limits, then puts hard ceilings outside the process where a bug cannot skip them.
Guide10 min read
Agents & multimodal
Long context in production: cost, latency and context rot
A 1M-token window is a ceiling, not a target. This guide covers what fills a context window, how vendors price long prompts, why recall degrades as context grows, and how to budget, trim and compact context in code.
Guide9 min read
Agents & multimodal
MCP and tool calling in production: failure modes and fixes
Tool calling moves the risk from what a model says to what your code does. This guide lists the ways MCP and function-calling integrations fail in production, the signal each leaves, and a wrapper that fixes them.
Guide10 min read
Agents & multimodal
Where the tokens go in a RAG pipeline
A RAG answer is billed in several places, and most of its input tokens are context you chose to send. This analysis breaks one answer into segments, shows which ones grow, and gives code to measure them per request.
Analysis8 min read
Agents & multimodal
Tracing an AI agent run across many model calls
An agent run is dozens of model calls, tool calls and retries. This guide shows how to give the run one identity, propagate it with W3C Trace Context, name each step, and read a run's cost without mistaking it for the whole bill.
Guide9 min read
Agents & multimodal
The AI agent infrastructure stack for production systems
Agents multiply requests, tools, and failure modes. The production stack needs bounded loops, durable state, model policy, cost controls, and traceable inference.
BriefUpdated 2 min read
Agents & multimodal
Production RAG infrastructure: beyond the vector database
The vector database is one box in a RAG system. Reliability depends on ingestion quality, retrieval policy, context construction, evaluation, and model routing.
BriefUpdated 2 min read
Agents & multimodal
One multimodal AI API for text, image, video, and audio
A unified API should standardize identity, policy, jobs, webhooks, observability and usage records — not pretend text streams and video generations behave the same.
BriefUpdated 2 min read