AgentPostmortem launches public registry to document and learn from AI agent failures

·3 views

A developer has built AgentPostmortem, a publicly searchable registry cataloguing real-world AI agent failures such as deleted databases, misdirected emails, and unexpected cloud bills. The platform, available at agentpostmortem.com, was created to address the lack of any structured, shared record of AI agent incidents — a gap that allows the same mistakes to repeat across teams. Submissions are structured with fields covering the agent involved, the prompt given, outcome, damage severity, and estimated financial cost, and can be made anonymously or with attribution. Each approved case is assigned a permanent identifier in the format APM-XXXX, enabling it to be cited and referenced over time. All submissions go through moderation and PII redaction before publication, with the project drawing inspiration from aviation safety boards and software postmortem culture.

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)

AI Shifts from Text Generation to Autonomous Action, Reshaping Developer Workflows

The AI landscape is undergoing a strategic shift, with models moving beyond text generation toward executing real-world tasks such as navigating interfaces, running end-to-end workflows, and operating development tools autonomously. GPT-6 Astra has drawn attention for its emphasis on 'computer use,' enabling agents to interact with actual operating systems using mouse and keyboard inputs. For frontend and product engineers, this evolution means AI can now act as a synthetic user capable of testing real UIs, CI pipelines, and design tools — not just assist with writing or code refactoring. On the business side, a two-tier pricing model is emerging: a standard tier with stronger data protections and a cheaper tier that requires users to allow their inputs to be used for model training, making tier selection a compliance and architecture decision for development teams. Analysts note a gap between highly polished demos and real-world benchmark performance, particularly around task completion speed, which remains a practical limitation for production use.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

New Pipeline Forces AI-Written Docs to Cite Source Files or Get Rejected

A proposed documentation workflow requires every AI-generated claim to be traceable to an existing repository file before it can be merged via pull request. The system distinguishes between facts a model may restate — such as API parameters from OpenAPI specs or commands from Makefiles — and promises that only humans can own, like stability guarantees or support windows. Unbound sentences, those lacking a link to a specific file, symbol, or architectural decision record, are treated as build failures rather than style issues. The pipeline runs four sequential stages: inventorying allowed source artifacts, extracting claims as structured records, binding each claim to a source, and rejecting the pull request if any claims remain unbound. The approach aims to prevent AI-drafted prose from inadvertently elevating beta features or undocumented behaviors into implied product commitments.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

BrowserSkill Lets AI Agents Use Your Logged-In Browser Without Hijacking Sessions

BrowserSkill, an open-source tool released by Tencent with 1.8k GitHub stars under the MIT license, allows AI agents to operate a user's existing browser instead of launching a new headless instance. The tool consists of two local components — a CLI daemon and a browser extension — enabling agents like Cursor, Claude Code, and Codex to drive a real, already-authenticated browser session. Because it reuses an existing browser with an established device history and valid cookies, it bypasses the fingerprinting and risk-control barriers that typically block automated tools. A dedicated Agent window keeps browser tasks visible and separate from the user's own work, while a built-in human-in-the-loop mechanism pauses the agent and requests manual input when it encounters captchas, logins, or sensitive confirmation dialogs. The project positions itself as complementary to headless automation tools like Playwright, handling tasks that require genuine login state rather than attempting to replicate or fake trusted sessions.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Lock Output File Hashes Before Refactoring Legacy Scripts, Dev Guide Advises

A software development guide published on DEV Community recommends capturing SHA-256 hashes of output files as a safety net before refactoring brownfield codebases. The approach involves running a script against fixture data, hashing the resulting file tree, and saving that hash manifest as a golden reference. On subsequent refactoring passes, the test suite re-runs the script and compares fresh output hashes against the saved manifest to confirm behavior has not changed. The method deliberately excludes volatile data such as timestamps, process IDs, and working-directory paths to keep the manifest stable across machines. A sample inventory CLI called stocktake.py, along with helper scripts hash_tree.py and test_stocktake_tree.py, is provided to illustrate the full workflow.

0 comments Read more at DEV Community