August 2026 Summaries

Originally inspired by Joel Spolsky’s 2000 checklist for assessing software-team effectiveness, the proposed “Shelley Test” updates those principles for an era in which coding agents and large language models play a central role in development. It argues that teams should use adversarial agentic code review rather than routine peer review, deploy continuously under LLM supervision, maintain trustworthy end-to-end tests, and provide agents with queryable observability data that combines system metrics and business context. The framework also emphasizes access to leading AI models, rapid merge queues and CI to avoid attention-draining delays, and infrastructure that lets teams quickly create and iterate on internal tools and agents. It encourages continual team discussion of workflows as practices evolve, and concludes that products increasingly need to be operable by coding agents through accessible APIs, reliable authentication, and clear machine-readable documentation.

After adopting a rule in February to stop writing code by hand, the author shifted from direct implementation to guiding AI agents through prompts, tools, testing, and environment design, finding that the change enabled more parallel work but created new coordination, validation, and security challenges. Early local multi-agent setups suffered from shared Git state, ports, processes, and dependencies, leading to isolated disposable VMs, automated environment provisioning, and botd, a remote, mobile-oriented system for managing agent tasks and preserving their conversations. The author allowed agents broad autonomy within isolated environments while restricting external access through read-only permissions, test-only write access, and credential proxies, emphasizing the risks posed by combining private data, untrusted inputs, and external communication. Agents were used not only for coding but also for customer-issue investigations, security testing, and deployment monitoring, though human judgment remained necessary to assess whether a technically correct change was worth shipping. The account argues for “agentic engineering,” in which humans define architecture, contracts, constraints, migrations, and validation before agents generate code, rather than accepting opaque AI-built systems through “vibe coding.” botd ultimately failed because its architecture had not been sufficiently understood, but its preserved SQLite data remained useful, reinforcing the conclusion that AI has changed the author’s work from iterating primarily on code to iterating on prompts, designs, and systems without eliminating the need for engineering responsibility.

exe.dev routes HTTP and SSH connections through globally distributed proxy servers that need current VM routing and access information without repeatedly querying a centralized database or replicating sensitive, irrelevant data everywhere. To balance performance and consistency, proxies cache data after an initial database lookup, while the system invalidates cached entries whenever relevant routing or sharing information changes. Its Go code uses a modified version of sqlc, which generates type-safe database functions and additionally records the tables and row-selection fields involved in reads and writes. Read operations accumulate dependency information through Go contexts, while committed writes invoke hooks that identify potentially affected cached data. A centralized writer process sends invalidation messages to proxies whose caches depended on modified rows, allowing them to conservatively discard related entries, such as all cached VM data for a given owner, and retrieve fresh information only when needed.

Frequent deployments reduce the risk and disruption associated with large, infrequent releases, but continuous delivery traditionally requires difficult automated systems to detect the many possible failures that can occur during rollout. At exe, an AI agent called Athena supports this process through the “exe-ops” Deployment Command Center, which evolved from shell scripts into a custom user interface rather than adopting a platform such as Spinnaker. Athena accesses Git repositories, metrics, and logs to assess whether deployments should proceed, select machines for rollout waves, pause or reject risky releases, escalate issues to people, and communicate updates through Slack. By reviewing code changes, analyzing logs, identifying unexpected problems, handling unusual issues, and suggesting improvements, the agent provides more consistent oversight than busy engineers may realistically be able to offer. This allows teams to deploy more often while reserving human attention for meaningful exceptions, making interventions smaller and less frequent.

Coding-agent users often fall between planning detailed specifications with an agent and rapidly iterating from short prompts, reflecting a longstanding product-development tension rather than a clear right approach. While upfront plans, PRDs, and slides can establish direction, prototypes often reveal the most valuable information: unexpected customer behavior, scaling failures, design changes, workarounds, contentious review issues, and meaningful metrics. Because agents can perform much of the experimentation, teams can ask them to build first and then surface the practical lessons and design implications, making prototypes a way to deepen and improve plans.

exe’s new Workload Identity Federation integration enables agents and workloads to access cloud resources through short-lived, narrowly scoped credentials rather than storing long-lived secrets in virtual machines. The approach builds on a model popularized by Kubernetes, where a workload receives a signed identity token from a trusted identity provider and exchanges it with a cloud provider for temporary access to an authorized service account. This reduces risks associated with leaked, untracked, and periodically rotated credential files while improving visibility into which workload is accessing which resources. exe’s implementation lets users create an Identity Federation integration, attach it to VM tags or individual VMs, and configure providers such as AWS or GCP to trust exe’s OIDC issuer. In a Google Cloud example, an agent requesting BigQuery access receives an exe identity token, exchanges it with Google’s Security Token Service for a federated token, impersonates an authorized service account through IAM, and uses the resulting temporary access token to query BigQuery.

The organization describes an internally built ecosystem of AI agents and tools that automate operational, engineering, and publishing tasks, including systematic security reviews focused on recent changes, alert investigations using logs, metrics, source code, and prior cases, and daily analysis of notable log trends. Other systems identify and propose fixes for flaky or slow CI tests, monitor deployments and rollouts against relevant code changes and system signals, provide an independently hosted status page, and send urgent phone alerts through PushOver rather than PagerDuty. Its blog has evolved from Git-based Markdown into a collaborative CMS with revision history, comments, embargoes, and AI-assisted importing, while UI tests can be expressed in natural-language descriptions that are converted into browser actions and potentially repaired by an LLM during CI. Reporter bots also summarize recent activity across commits and support threads, with the organization emphasizing isolation and restricted tool permissions to reduce risks from agents processing untrusted content alongside private data and external communications.

Following Bending Spoons’ reported $1.28 billion acquisition of Airtable, a former Airtable employee argues that advances in LLMs with tool use may reduce the need for traditional no-code platforms. The piece traces how spreadsheets, low-code tools, and SaaS enabled business users to create applications when IT procurement and infrastructure barriers limited bespoke software, while praising Airtable’s approachable database modeling experience. It contends that coding agents can now help non-engineers build, deploy, and iterate on custom web applications directly on Linux-based infrastructure using open-source stacks, with less platform lock-in than proprietary no-code services. The author promotes exe.dev as a service offering secure Linux virtual machines, integrated coding agents, sharing controls, automation support, integrations, and plans starting at $20 per month, while suggesting that users can migrate existing spreadsheets or low-code systems through agent-assisted conversion and add more formal development practices only as applications become business-critical.

exe.dev integrations keep secrets outside virtual machines by using a proxy to inject credentials through mechanisms ranging from HTTP headers to custom integrations. Its updated GitHub integration now provides a single URL format, https://github.int.exe.xyz/OWNER/REPO, that automatically selects the appropriate authorized integration for any accessible repository, including team integrations, while retaining compatibility with existing Git remotes. Users can also configure GitHub integrations as read-only, limiting both Git access to fetching rather than pushing and GitHub API access accordingly.

At exe, bots, agents, and LLM-generated outputs are treated as computer programs that assist with organizing, monitoring, and maintaining the community while following established etiquette. Bots are kept out of human discussion spaces because they cannot assess the attention cost of their messages, and people are not expected to read activity in bot-focused spaces. Humans are encouraged to write in their own words, with any pasted bot output clearly labeled, commonly through block quotes or code blocks. Each bot post identifies both its runtime location and source code, while bots may be given names for ease of reference but are consistently referred to with it/its pronouns.

Shelley Tokens can now be automatically purchased when a user’s balance reaches a configurable floor, reducing the need for manual credit purchases. Users can set the trigger threshold, the number of tokens bought per purchase, and a monthly spending cap intended to prevent uncontrolled costs, with controls available through SSH billing commands or a redesigned dashboard page. Auto-purchasing is also available to Team plan users, although it currently charges each individual’s personal payment method while shared token and payment solutions are still being developed.

Exe.dev offers a robust solution for remote terminal management by utilizing two Ghostty instances to ensure seamless functionality across devices and networks. One instance manages the terminal rendering in the app or browser, while the other provides scrollback on reconnection within the session manager on the VM. The platform's key features include automatic reconnection during network roaming, session management for resuming activities across devices, scrollback capability, and modern terminal support with features like colors. Exe-scroll acts as the session manager process, allowing users to reconnect to their sessions via HTTPS or SSH, with the platform supporting a single session with scrollback. While many existing tools like Mosh, Eternal Terminal, and terminal multiplexers like tmux and screen provide similar functionalities, exe-scroll fills a specific niche by offering a simplified single-session management solution. The text acknowledges the influence of prior art in the field but emphasizes exe.dev's unique positioning in providing a convenient and reliable terminal experience tailored for its users.

The text discusses the evolving landscape of software personalization, highlighting a shift from traditional customization methods reliant on configuration files and plugins to a more dynamic, agent-driven approach. This transformation is facilitated by the use of open-source agents that allow users to modify and personalize software efficiently without extensive programming knowledge. The article illustrates this concept through a personal project that integrates a tool called "meat.dev" into an agent named Shelley, streamlining workflow processes by automating code review tasks and minimizing unnecessary details. It underscores the benefits of personalizable software for individuals and small teams, suggesting that the ease of modifying software with agents negates the need for extensive configuration systems, allowing for more tailored and efficient solutions. The narrative contrasts open-source agents like Shelley and Pi, which offer seamless personalization, with closed-source software like Claude Code, which limits customization to predefined hooks, advocating for a future where software is inherently designed for personalization through accessible source code.