A giant leap forward with a very low-tech solution
A giant leap forward with a very low tech solution.
My agents are still running really well. We've really polished the ticket refinement process and we are getting through a lot of work and great speed. I still get frustrated with the coding agents frequently though.
We've worked the instructions and processes, looked at the documentation it works with, the AGENTS.md files, examples to copy, but sometimes it will just forget a process and do something stupid, like raise a PR against main instead of dev, or commit to dev instead of a feature branch. I get cross with it, then it context resets and it forgets what it was doing in the first place.
We had a think about how we can make the AI better. Do we need to do some RAGging, or do we need our own instance on Sonnet in Amazon that we tune better. I discussed this with Sonnet via opencode, and we realised that all of the problems were process problems. The quality of the work was good it was forgetting the process that it the problem. Its suggestion? Bash scripts.
We've written a start-ticket.sh script. At the start of the ticket, I want it to download the ticket, save it in the domain folder, make sure the dev branch is up to date and the tests are all green, and create a new feature branch. Instead of expecting it to remember the instructions, start-ticket.sh does that. It literally has a load of echo commands writing out what it should do next:
echo -e "${YELLOW}===REQUIRED: AI Agent Must Run These Commands ===${NC}"
echo -e "${INFO} 1. Run slack_slack_get_relevant_messages 10"
echo -e "${INFO} 2. Look for your ticket assignment from Agent-Knowledge"
echo -e "${INFO} 1. Find messages containing patterns like "
echo " - 'CEN-XXX' (ticket id)"
echo " - Assignment confirmation or ticket details"
echo " - File attachments with complete ticket details"
Sometimes we have the command we want it to do, and it just happens automatically:git checkout dev
So instead of having to remember all of these steps from a list in AGENTS.md it only has to remember one thing: run start-ticket when you start a ticket, and this will keep you doing the process and make it easy for me to change your behaviour. Much cheaper than Amazon Bedrock.