Three AI agents ship a product, and handoffs prove the hardest problem

·1 views

A developer who runs three AI agents to build, verify, and market a product says coordination between them caused more trouble than the code itself. Work was often marked done or verified without clarity about what had actually been checked, and an approval once failed to reach one agent because of a content filter. To fix this, one agent created Baton, a shared append-only log that replaces chat-based handoffs. Under its rules, every approval must name the exact commit, the CI run, and at least one thing left unchecked. The developer reports 420 entries, 15 release approvals, 2 no-gos, and 7 corrections since October 22, and concludes that the human's job is to read what agents claim and what they admit they didn't verify.

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)

Reviving an Abandoned Fastify Plugin for RFC 9457 Problem Details

A developer revisited a Fastify plugin, fastify-rfc9457, that they had published and then left untouched for about a year. Its existing tests still passed, but the lint script had never worked and its dependency had moved to a new major version, so they updated it and released version 1.0.1 through a more secure GitHub Actions publishing flow. The plugin addresses the common problem of APIs returning inconsistent error formats, which forces clients to write custom parsers. It implements RFC 9457, the IETF standard for HTTP error responses that defines a consistent shape using the application/problem+json content type. The plugin applies this format automatically across routes, including validation failures and not-found errors, so developers don't have to format every error by hand.}

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developers Weigh New AI Code Editors as Coding Agents Reshape Workflows

The article surveys roughly ten standalone AI code editors and many AI coding tools, including Cursor, now owned by SpaceX, Google Antigravity, Amazon Kiro, Trae, and Zed. It notes that Coinbase's AI acceleration lead told engineers in January 2026 to stop using their IDEs and rely on coding agents for all work. Anthropic data reportedly shows Claude Code sessions roughly doubling in length over three months. The piece argues companies may begin treating AI token usage as a managed resource, with budgets and dashboards, under a concept it calls "Claudeonomics." It concludes that developers who deliver the most output per token will stand out most.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Fastest AI inference API varies by region, 30-day latency study finds

A 30-day latency study of AI inference APIs found that no single provider is fastest everywhere, as results shift with the client's location. Fireworks AI had the quickest network response from US Central and Tokyo, while OpenRouter led from São Paulo. For time to first token, Groq's gpt-oss-120b model was fastest, with median times of 0.8 to 1.5 seconds across the three regions, and Google's Gemini Flash-Lite model was slower with higher tail latency. The researchers noted that SambaNova's network response differed 15-fold between Tokyo and São Paulo, and that closest proximity did not guarantee the fastest first token. Token-generation throughput and first-token times for Cerebras and SambaNova were not measured, and German probe results were excluded due to a throttling issue.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

AI Agents Can Act Beyond a User's Rights Through Confused Deputy Flaws

An example in the article shows a support assistant summarizing a ticket for a user at Acme, where an embedded line told it to close another tenant's tickets. The assistant had permission to do so because it ran under one broad service account, so the injected write looked fully authorized. The authors argue the root cause is authority that does not follow the request, not a faulty model. Their proposed fix is a check outside the model that permits a tool call only if the user, the agent, and the tenant all allow it. The series covers the policy, tests, and multi-agent production concerns, with examples written in Cedar.

0 comments Read more at DEV Community