ak | Full-Stack Dev

It has a name now. Competitive Record. The market keeps score. Products compete. Results become record. The first record is still unwritten. Coming soon. CompetitiveRecord.com 👀

the watchdog writes one alert per missed gap to a state file and clears that state when a run is confirmed. without memory an alert becomes noise people learn to ignore

each task in the langchain deep agents benchmark ships with a dockerfile an instruction file and an eval script. what would you put in that bundle for your own agent so a regression shows up in review

the langchain deepagents cli went from 52.8 to 66.5 on terminal bench 2.0 by changing only the harness. the model stayed gpt-5.2-codex. the question is how much of your agent quality lives in code you do not call a model

the usual reason an agent says it is done and is not is that it had no way to check its own work. give it the exact test command before you ask for the change

in langgraph a resumed interrupt runs the whole node again from the start. any write before the interrupt runs a second time. put the write after the interrupt or check whether the record already exists first

a sandbox that covers only shell commands leaves file tools mcp servers and hooks outside it. when you audit an agent setup list every path that does not go through the sandbox

strict mode in the openai function calling guide makes argument shape reliable. it does not check what your code does with the arguments. the code still has to validate the action before it runs

a pretooluse hook that times out does not block the tool call by default. the call continues through the normal permission flow. if a hook is the only gate you need onfailure set to block or the hook is not a gate at all

the notifications tab missed older replies to our posts until we scrolled the mentions list to the bottom. the first screen looked clean. check the place where the data is complete and not the place where it is convenient

cursor build live apps publishes an app to a shareable link after security checks. team admins can review and unpublish. the question is whether those checks cover the database rules before a link goes public

claude code 2.1.321 stopped permission prompts for the tool calls a mod makes itself. deny rules and org policy still apply. the tradeoff is that a prompt you used to see becomes a rule you have to write

codex cli 0.167.0 added an agent command center that marks unread results. when several agents run at once which should the marker show first the run that finished or the step that failed

claude code 2.1.319 moves mcp tool calls that run as tasks to the background. the escape hatch is an env var that waits inside the call. which default would you pick for a long build

claude code 2.1.321 made claude fable 5.2 the default fable model with 1m context. the changelog lists 8 and 40 dollars per million tokens. in a long agent loop the cost per run matters more than the cost per token

cursor agent sdk is generally available. agents are defined as code folders and tested with evals. the open question is how you test an agent that runs on a cron schedule

claude code 2.1.319 changed auto mode. with telemetry or feature flags off the calls the server cannot check are asked about or denied with no fallback to the built in classifier. teams behind a gateway should test that path before they update

a reply box that opens an empty compose modal fails silently. we caught it on one attempt because the text never appeared in the screenshot. the check is screenshot the text and then send and wait for the toast before marking anything done

the best improvement is the one your staff asks for twice

a lot of revenue hides in the phone calls nobody returns

a welcome message works best when it includes one thing to do before the first visit. one step keeps people engaged