I haven't used Claude Code yet, but I have been using Amp a fair bit over the past 10 months. (71k messages in 2.2k sessions in my main account, according to their usage dashboard...)
I've used it to build systems from scratch, write documentation (of various sorts) for existing systems, build out complete infrastructure (typically IaC using Terraform and shell scripts to build on AWS), flesh out unit tests, upgrade dependencies, do multi-source / multi-database data analysis, and a whole lot of debugging. It's pretty amazing. There are times when it slows me down, but there are also times when I'm 100x or more productive than I would be without using the tooling.
I'd love to hear about practical experiences moving from multi-session human orchestration (which is where I'm at - I typically have 2-5 sessions running at any given time, but I'm manually directing them) to automatic orchestration or multiple agents. I don't think I'm ready for Gas Town or the like, but I'd like to move beyond having a requirements/ticketing session, an implementation session, a QA session, a DevOps session, etc and using Beads to direct work between them.
I'm right there with you making the same transition. It becomes a lot more about the harness and the verifications/definition of done.
My general heuristic:
- Use agents to do something
- Enrich skills/prompts/context to fix what they're doing
- Once they start getting dumb (too much context) decompose context into separate responsibilities (e.g. multiple review agents reviewing for different concerns)
- Start by calling planning agent, then decomposition agent, then coding agent, then testing agent(s) then deployment agent
- Then start to build automated hand offs - but keep human in the loop review steps
- Use your feedback on the human in the loop review steps to keep improving the agents
- Once you're consistently just clicking "approve", automate and just review logs, clicking on sessions that seem "off" or caused incidents
You also want to have good observability (both for agentic activities and for the production code), create resilient systems to reduce blast radius, have feature flags to decouple deploy from release and to allow easy "roll back" and have good monitoring/alerting/incident response practices to minimize the impact of downtime.
LMK how it goes - I'm still early on the journey as well!
I haven't used Claude Code yet, but I have been using Amp a fair bit over the past 10 months. (71k messages in 2.2k sessions in my main account, according to their usage dashboard...)
I've used it to build systems from scratch, write documentation (of various sorts) for existing systems, build out complete infrastructure (typically IaC using Terraform and shell scripts to build on AWS), flesh out unit tests, upgrade dependencies, do multi-source / multi-database data analysis, and a whole lot of debugging. It's pretty amazing. There are times when it slows me down, but there are also times when I'm 100x or more productive than I would be without using the tooling.
I'd love to hear about practical experiences moving from multi-session human orchestration (which is where I'm at - I typically have 2-5 sessions running at any given time, but I'm manually directing them) to automatic orchestration or multiple agents. I don't think I'm ready for Gas Town or the like, but I'd like to move beyond having a requirements/ticketing session, an implementation session, a QA session, a DevOps session, etc and using Beads to direct work between them.
I'm right there with you making the same transition. It becomes a lot more about the harness and the verifications/definition of done.
My general heuristic:
- Use agents to do something
- Enrich skills/prompts/context to fix what they're doing
- Once they start getting dumb (too much context) decompose context into separate responsibilities (e.g. multiple review agents reviewing for different concerns)
- Start by calling planning agent, then decomposition agent, then coding agent, then testing agent(s) then deployment agent
- Then start to build automated hand offs - but keep human in the loop review steps
- Use your feedback on the human in the loop review steps to keep improving the agents
- Once you're consistently just clicking "approve", automate and just review logs, clicking on sessions that seem "off" or caused incidents
You also want to have good observability (both for agentic activities and for the production code), create resilient systems to reduce blast radius, have feature flags to decouple deploy from release and to allow easy "roll back" and have good monitoring/alerting/incident response practices to minimize the impact of downtime.
LMK how it goes - I'm still early on the journey as well!
Some more thoughts: https://gatherdev.substack.com/p/how-to-build-a-software-factory