Action Context Window Connecting Pickaxes

General
1

Quick question.

When connecting Pickaxe’s together, how much of the memory/context is carried over?

Example - let’s say I’ve got a “Sales copy Pickaxe” (Parent)

That pickaxe is than connected to a different pickaxe for specific functions (sales pages, order bumps, emails, ads, etc). (Sibling)

Is the frame that you should be then brain dumping context with the parent pickaxe and then that carries over to the siblings?

Or would you build it so that each sibling has custom instructions to ask a few questions and gather more data/detail as intended ?

1 Like
2

Hey @zachansen Short answer: It depends on what you explicitly hand off. Directly chaining Pickaxes won’t magically transfer “everything.” You control the carry-over via the trigger prompt and any variables you pass forward.

What actually carries over

  • User input you forward >> whatever you include in the next Pickaxe’s trigger prompt (or mapped vars)
  • Named variables >> values you store/emit and then reference downstream
  • Nothing else by default >> the prior chat’s full history/KB isn’t auto-available unless you deliberately pass it or re-query it

How to make chaining reliable

  • Define a handoff contract >> a tiny JSON envelope (e.g., goal, key facts, constraints, links) the first Pickaxe outputs and the second Pickaxe reads
  • Be explicit in both system prompts >> “Only use the provided handoff; do not infer missing details; ask for clarification if field X is empty”
  • Keep context lean >> pass just the essentials needed to answer the next question; link to docs instead of dumping big blobs

Prompt design that helps

  • Trigger prompts mirror each other >> same field names, same expectations
  • Validation step >> first Pickaxe confirms the handoff fields are populated before calling the second
  • Guard against inference >> “If information isn’t present in handoff, ask, don’t invent”

When you’ll hit limits

  • Large handoffs >> long summaries can get truncated; prefer pointers (doc IDs/URLs) and fetch on demand
  • Ambiguous asks >> if the next Pickaxe’s trigger is vague, it will infer; tighten the schema or add a quick clarifying question

If you need deterministic handoff

  • Use a Make webhook / n8n layer to carry a small, structured payload between tools (goal, context, IDs) and re-pull any heavy KB in the second step.

I go deep on this in PPA, including templates that auto-generate both system and trigger prompts so the handoff is clean. If you want, I can share a minimal handoff schema you can drop in.

3 Likes
3

Joining now, thanks for the reply!

Quick question - I’m finding through my testing of chaining that artifacts are not actually getting built. So, as an example, if I trigger the parent tool to trigger the sibling it just constantly refreshes and doesn’t actually build the artifact :thinking:

1 Like
4

Are you asking the sub-agent to build an artifact via the master agent? If yes, that won’t work.

The master agent has to be the agent with Copilot connected in order for the artifact to output. Does that make sense?

2 Likes
5

Yes that’s exactly what I did and it worked…was just very slow.

Thinking I’m going to turn the artifact “off" for now because it takes forever to output

1 Like
6

Try changing the model. Try Claude 4.5 Sonnet or Grok 4

1 Like
7

Thanks for sharing that, Zach! Pickaxe works as a bridge between you and the AI model, so if the model takes longer, it reflects on Pickaxe too. Complex prompts, too many actions, or slower models can also add time. Try switching models to see which one responds faster since each behaves a little differently.