Agent Rails #5

New Issue

2026-09-21T12:43:36+02:00

g4borg commented

2026-09-21 12:43:36 +02:00

-> It seems not enough to tell the agents to not use Explore or Plan Agents only, we also need to help codify the agent cycle better

each agent we use that should produce a step (research, report, design, plan, ...) should have the ability to WRITE. Explore and Plan agents are READ ONLY in claude. they become useless for igor and all its subagents in 99% of the time.

-> Codify the CYCLE also a central concept: the cycle means that we have certain goals, and produce proper designs, plans and implementations from it.

brainstorms and alignment: usually not part of the cycle, but happens before it.
research: sometimes we need to research open questions. most of the time, the research already decides most decisions.
design: usually inside the cycle, design with open questions should still be folded by the igor, by selecting whatever the most likely decision is. it is important to note that some open questions might become blockers, if the consequences of implementation are extremely different depending on the answer. taking the best decision, and flagging it in the end is another possibility how to move on from design "for now".
plan: this is the entry to the loop actually -> an agent writes a proper plan, sometimes multiple plans.
plans land in the athanor, given that works sequentially, they are queued to be added if added later.
once athanor implements the plan, we want to review the code. if any issue arises that is medium or higher, we collect these, and write a new plan. this is why its called the cycle: the review decides, if we go back into an earlier step and iterate over the result. most of the time this is a plan. in very few occasions, a review might open an entirely new problem space, that requires research, maybe a design step or implementation. as long as this is not really major (like an entirely new crate, or suddenly changing core mechanics of a part of existing code we did not plan to touch in a major way) we loop here until we are done.

-> "Athanor is in the oven"

I hear igor saying that a few times, showing that this is not clarified well. Athanor IS the oven. putting something into the oven should mean to let athanor cook up the plans into code, or whatever outcome we want (sometimes agents also used athanor to do multiple plans and designs, sometimes even to coordinate research-trees)

-> It seems not enough to tell the agents to not use Explore or Plan Agents only, we also need to help codify the agent cycle better - each agent we use that should produce a step (research, report, design, plan, ...) should have the ability to WRITE. Explore and Plan agents are READ ONLY in claude. they become useless for igor and all its subagents in 99% of the time. -> Codify the CYCLE also a central concept: the cycle means that we have certain goals, and produce proper designs, plans and implementations from it. - brainstorms and alignment: usually not part of the cycle, but happens before it. - research: sometimes we need to research open questions. most of the time, the research already decides most decisions. - design: usually inside the cycle, design with open questions should still be folded by the igor, by selecting whatever the most likely decision is. it is important to note that some open questions might become blockers, if the consequences of implementation are extremely different depending on the answer. taking the best decision, and flagging it in the end is another possibility how to move on from design "for now". - plan: this is the entry to the loop actually -> an agent writes a proper plan, sometimes multiple plans. - plans land in the athanor, given that works sequentially, they are queued to be added if added later. - once athanor implements the plan, we want to review the code. if any issue arises that is medium or higher, we collect these, and write a new plan. this is why its called the cycle: the review decides, if we go back into an earlier step and iterate over the result. most of the time this is a plan. in very few occasions, a review might open an entirely new problem space, that requires research, maybe a design step or implementation. as long as this is not really major (like an entirely new crate, or suddenly changing core mechanics of a part of existing code we did not plan to touch in a major way) we loop here until we are done. -> "Athanor is in the oven" I hear igor saying that a few times, showing that this is not clarified well. Athanor IS the oven. putting something into the oven should mean to let athanor cook up the plans into code, or whatever outcome we want (sometimes agents also used athanor to do multiple plans and designs, sometimes even to coordinate research-trees)

g4borg referenced this issue from a commit

2026-09-22 13:57:57 +02:00

🤖 Fix agent write capabilities, Athanor metaphor, and delegation guidance

g4borg referenced this issue from a commit

2026-09-22 13:57:57 +02:00

📝 Add development cycle design and session artifacts for issue #5

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: dirigence/reliquary#5