Indeed, you can become the glue for your AI agents. :) I see two ways out: one agent owns all the context, which feels like overkill, or specialized agents can talk to each other. Mine need a normal channel where they can ask each other for an update or more context, instead of using me for every handoff. Then they can become the glue for each other.
Yes — this is exactly the direction I’m arriving at too.
And I think there may be one step beyond agents simply talking to each other: giving them explicit shared state, so they don’t have to keep asking each other what happened, what is missing, and what should happen next.
Maybe the goal isn’t to automate every handoff. Sometimes it’s to design the company so the handoff isn’t needed in the first place.
This is actually where I’m going in the next piece I’m publishing shortly. :)
I started there too. Shared state is useful, but I don’t think it is enough. I began with shared context, and it kept growing. The agents became less efficient, started prioritizing one part of it over another, and sometimes old context even poisoned later decisions. Now I’m separating their contexts again. I think they need both: a minimal shared state for coordination, and a way to ask each other for the specific context they need.
That makes sense — and I think the distinction may be between shared context and shared state.
I wouldn’t want every agent to carry or read the whole shared context either. I’m thinking more about partitioned state with clear ownership: one authoritative writer for a given piece of state, many possible readers, and each agent retrieving only the minimum it needs for the task at hand.
Then agent-to-agent communication can handle the specific context that isn’t worth keeping in shared state.
So maybe the architecture is actually a combination of both: minimal structured shared state for coordination, plus targeted conversations when more context is needed.
Indeed, you can become the glue for your AI agents. :) I see two ways out: one agent owns all the context, which feels like overkill, or specialized agents can talk to each other. Mine need a normal channel where they can ask each other for an update or more context, instead of using me for every handoff. Then they can become the glue for each other.
Yes — this is exactly the direction I’m arriving at too.
And I think there may be one step beyond agents simply talking to each other: giving them explicit shared state, so they don’t have to keep asking each other what happened, what is missing, and what should happen next.
Maybe the goal isn’t to automate every handoff. Sometimes it’s to design the company so the handoff isn’t needed in the first place.
This is actually where I’m going in the next piece I’m publishing shortly. :)
I started there too. Shared state is useful, but I don’t think it is enough. I began with shared context, and it kept growing. The agents became less efficient, started prioritizing one part of it over another, and sometimes old context even poisoned later decisions. Now I’m separating their contexts again. I think they need both: a minimal shared state for coordination, and a way to ask each other for the specific context they need.
That makes sense — and I think the distinction may be between shared context and shared state.
I wouldn’t want every agent to carry or read the whole shared context either. I’m thinking more about partitioned state with clear ownership: one authoritative writer for a given piece of state, many possible readers, and each agent retrieving only the minimum it needs for the task at hand.
Then agent-to-agent communication can handle the specific context that isn’t worth keeping in shared state.
So maybe the architecture is actually a combination of both: minimal structured shared state for coordination, plus targeted conversations when more context is needed.