MCP Goes Stateless in 2026, But AI Agents Still Require Durable State

·6 views

The July 2026 Model Context Protocol specification introduced a stateless protocol core, removing persistent session identifiers so that requests can be handled independently across multiple servers. While this simplifies infrastructure scaling using standard HTTP load balancing, developers risk mistaking stateless transport for stateless workflows. AI agents handling multi-step tasks — such as refund approvals that may pause for 30 minutes or more — must preserve workflow state across process restarts, deployments, and server switches. Storing agent state only in process memory causes failures in distributed environments, where a resumed workflow may land on a different server instance that has no record of the original run. The recommended approach is to persist workflow state externally using tools like PostgreSQL, Redis, or durable execution runtimes, making individual server instances fully replaceable.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

OSLabs Releases Open-Source CLI PennyWyze to Audit Claude API Costs

Three engineers at OSLabs have released PennyWyze, a free open-source command-line tool that helps developers find the cheapest Claude model tier that still meets their accuracy requirements. Users provide their production prompt and a labeled dataset of real examples, and the tool runs these against Claude's Opus, Sonnet, and Haiku tiers using the live Anthropic API. PennyWyze grades each model's responses against expected answers and calculates actual monthly cost projections based on real token counts. In one example audit, the tool identified that switching from Opus to Haiku could save roughly $180 per month with no loss in accuracy. Currently, PennyWyze supports exact-match grading, making it best suited for classification, extraction, and routing tasks rather than open-ended generation.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Free API tracks OpenAI Codex global usage-limit resets with 24/48h forecasts

A developer has launched codex-reset.com, a free tool that monitors and records OpenAI's informal global Codex usage-limit resets, which occur occasionally after incidents or milestones. The service provides a public JSON API requiring no key, offering probability forecasts for a reset occurring within the next 24 or 48 hours alongside a dated timeline of past resets with source links. It also tracks live Codex service status across multiple surfaces, helping users distinguish outages from quota exhaustion. The data is additionally available as a read-only MCP server that can be connected directly to Codex or Claude Code without installation or authentication. The project is free to use, with the only condition being visible attribution when the data is displayed to others.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

MCP vs CLI: The Right Choice Depends on Who Controls the Next Step

When building tools that perform actions like publishing a website, developers face a choice between a Command Line Interface (CLI) and a Model Context Protocol (MCP) server. The key deciding factor is not whether the user is human or an AI agent, but rather who determines the next step in the workflow and where the data resides. CLIs are better suited for fixed, repeatable sequences such as automated deployment jobs or local build commands, while MCP tools shine when an AI assistant needs to discover and invoke operations dynamically within a conversation. Transport context also matters: a local MCP server can read a directory path directly, whereas a remote MCP server over HTTP interprets the same path on the server side, not the user's machine. Both interfaces can share a single underlying operation, with the CLI managing flags and exit codes while the MCP adapter handles tool definitions and structured results.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Add AI Agent Billing Without Disrupting Its Workflow Loop

When a subscription gateway is placed between an AI agent and its model provider, it can mistakenly interrupt multi-step tasks by treating each model request as a separate billable turn. The core problem is that billing layers can inadvertently take over orchestration decisions that should remain with the agent's local host. A properly designed billing gateway should only handle authentication, allowance checks, request forwarding, and usage accounting — not control the flow of the agent's tool loop or conversation context. Developers are advised to assign distinct identifiers to user turns, model operations, and business invocations to enable accurate metering and error recovery without disrupting execution. Keeping orchestration deliberately separated from billing ensures that paying for model access does not silently grant a second system control over the user's task.

0 comments Read more at DEV Community