Four Design Lessons from Building a Block-Editing Language for AI Agents

·3 views

Developers building IWE, an AI agent memory system, found that requiring agents to re-emit entire documents for small edits caused a 9% structural malformation rate in their first benchmark round. To fix this, they created a block-level editing language that lets agents target specific document nodes—headers, paragraphs, tables—without rewriting whole files. The system introduces 'expect guards' that force operations to fail loudly if the number of matched targets differs from what the agent anticipated, reducing silent errors. Document keys must be derived deterministically from metadata, making duplicate entries a structural impossibility rather than a prompt-discipline problem. These design choices reflect a broader principle: for AI agents, safety constraints should be enforced by the engine, not left to instruction-following.

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)

How to Automate Google Search Console Indexing Using GitHub Actions

Manual URL submission to Google Search Console is impractical for large or frequently updated websites, prompting developers to seek automated alternatives. A command-line tool called gsc-indexer can be integrated into a GitHub Actions CI/CD pipeline to programmatically handle URL inspection and indexing requests. The workflow runs on a daily schedule, fetches URLs from a sitemap, and submits them to GSC using a Google Cloud Service Account for authentication. Credentials are stored securely as GitHub Secrets to prevent accidental exposure of sensitive keys. The pipeline also supports error monitoring by emitting JSON results and non-zero exit codes, enabling teams to trigger alerts when critical pages fail to get indexed.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer corrects MCP migration tool that confidently gave wrong upgrade advice

A developer building a compliance checker for the Model Context Protocol's major 2026-07-28 breaking changes discovered their own tool was giving incorrect migration guidance. The checker's rule MCP007 advised users to upgrade to @modelcontextprotocol/sdk version 2.x, a package that does not exist and never will, stopping at version 1.30.0. The MCP v2 release actually ships under entirely new package names, including @modelcontextprotocol/server, @modelcontextprotocol/client, and @modelcontextprotocol/core, among others. The developer corrected the rule to detect the presence of the old v1 package rather than comparing version numbers, and added automated tests to prevent the same mistake from recurring. The 2026-07-28 MCP revision removes stateful sessions, replaces server-initiated requests with multi-round-trip requests, and makes server discovery mandatory.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Flutter Devs Can Fix Riverpod Family Provider Stale Cache Issues Using Signals

A developer raised a common Flutter architectural problem on Stack Overflow where Riverpod 3 family providers storing parameterized queries in isolated silos caused stale and inconsistent data across UI components. When a shared entity like a task appeared in multiple filtered lists, updating it in one provider left other providers showing outdated information. The root cause applies broadly to BLoC, Redux, and Provider patterns, not just Riverpod, as parameterized fetchers act as disconnected local state containers. A proposed solution replaces these silos with a normalized entity store using mapSignal, where all entities live in a single key-value map and filtered views are derived as computed signals. This approach eliminates redundant network calls, stale relational data, and complex multi-provider invalidation logic by ensuring any single entity update automatically propagates to all dependent UI projections.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why Payout Systems Must Treat 'Unknown' as a Distinct State, Not a Failure

Modern payout infrastructure operates across multiple banking layers — including settlement queues, fraud engines, and batch processors — meaning a transaction can still be in progress even when an API call appears to have failed. Engineers often instinctively mark timed-out or unresponsive transactions as failures, but in payment systems this reflex can trigger duplicate payouts and ledger inconsistencies. Unlike typical software, payment rails introduce a third transaction state — 'Unknown' — where the final outcome cannot yet be determined. Forcing an unknown state into a failure state is a common source of production incidents, such as cases where a retry succeeds after the original payout had already cleared. The core architectural principle for resilient payout platforms is to never assume failure without definitive proof, keeping uncertain transactions in a pending state until reconciliation data confirms the outcome.

0 comments Read more at DEV Community