Love the harness diagram! For Memory & Context., I wonder if those eventually need to become two separate layers.
In my experience, context helps the agent make the current decision: account state, ICP, pricing, objections, recent calls. Memory preserves what the system learned from prior decisions: why a rule changed, what evidence caused it, and whether the change actually improved the outcome.
Once reps create their own skills and automations, that distinction feels more important. Otherwise the harness can accumulate more information without necessarily getting better at deciding what should persist. Curious how you and Kyle are thinking about that separation.
You're right, there is an important distinction, although both sometimes source from similar concepts, that's why we put them together. Context is what the agent needs for today's call and is case-specific. Memory is different: it's what got kept, why it got kept, and whether keeping it actually helped. Without that provenance, memory just becomes a bigger pile of old context.
The "did you build it on purpose, or is it a combo of prompts" line is the whole game. The teams I've watched get real leverage treat the harness as a context asset they curate — the lost-deal reasons, the pricing exceptions, the good discovery calls — and version it deliberately, rather than letting it accrete inside prompt strings no one can audit. Your build-vs-buy split maps cleanly onto that: buy the retrieval plumbing, build the context curation, because that's the part no vendor can know for you. Curious which of your six audit questions best predicts whether a team actually owns its harness versus just having one.
Worth noting the failure mode on the other side too: I've watched teams build "AI systems" so comprehensive that reps spend more time feeding context in than they save on output. The harnesses that actually get used tend to do one narrow thing extremely well, like call prep or objection handling, rather than trying to be a full GTM co-pilot. What's the narrowest version of this you'd recommend starting with?
Love the harness diagram! For Memory & Context., I wonder if those eventually need to become two separate layers.
In my experience, context helps the agent make the current decision: account state, ICP, pricing, objections, recent calls. Memory preserves what the system learned from prior decisions: why a rule changed, what evidence caused it, and whether the change actually improved the outcome.
Once reps create their own skills and automations, that distinction feels more important. Otherwise the harness can accumulate more information without necessarily getting better at deciding what should persist. Curious how you and Kyle are thinking about that separation.
You're right, there is an important distinction, although both sometimes source from similar concepts, that's why we put them together. Context is what the agent needs for today's call and is case-specific. Memory is different: it's what got kept, why it got kept, and whether keeping it actually helped. Without that provenance, memory just becomes a bigger pile of old context.
The "did you build it on purpose, or is it a combo of prompts" line is the whole game. The teams I've watched get real leverage treat the harness as a context asset they curate — the lost-deal reasons, the pricing exceptions, the good discovery calls — and version it deliberately, rather than letting it accrete inside prompt strings no one can audit. Your build-vs-buy split maps cleanly onto that: buy the retrieval plumbing, build the context curation, because that's the part no vendor can know for you. Curious which of your six audit questions best predicts whether a team actually owns its harness versus just having one.
Worth noting the failure mode on the other side too: I've watched teams build "AI systems" so comprehensive that reps spend more time feeding context in than they save on output. The harnesses that actually get used tend to do one narrow thing extremely well, like call prep or objection handling, rather than trying to be a full GTM co-pilot. What's the narrowest version of this you'd recommend starting with?
banger! nice work maja.
and thanks for mentioning my piece. was a blast to write that alongside George.
+ this piece may also be relevant for folks: https://www.thesignal.club/p/build-vs-buy-in-the-ai-era