Shipping Agents with CTO Henrik Feldt
Something shifted recently. AI went from just writing code to carrying out multi-day tasks autonomously. You can now tell your agent to write code, test it, fix errors and iterate. All while you sleep.
Henrik Feldt knows how to deploy agents like this as both the CTO of CopeCart & as the solo-founder at Aient.
Last month, his agents consumed 327 million tokens — $19,000 at standard API pricing. He paid $600.
That 32× gap is venture capital subsidizing your development. It will not last. But the architectural decisions you make today will follow you.
Henrik will pull back the curtain on what agentic engineering actually looks like — the parts that work, the parts that surprise you, and what will matter once the rules change:
The harness — the thing everyone ignores
The model is not what determines quality or cost. The harness is: the sandbox, the prompts, the context management, the guardrails. Henrik ran 94 agent restarts and 4.6 million tokens through a single optimization loop that produced nothing — until the harness changed. This is where the real engineering happens.
The subsidy math
The big AI companies are currently paying for much of your experimentation. When the gap closes, which decisions survive?
Budget and attention when everything is suddenly possible
When an agent can iterate on a problem for a week, the constraint is no longer capability — it's judgment. What do you give the agent, what do you keep, and how do you decide?
Trust and the merge button
How autonomous is too autonomous? What does it look like when a system makes decisions in production — and where is the line you don't cross?
Henrik has spent two decades starting companies and writing the software that keeps them running. This isn't a session about what AI might do. It's a conversation with someone building it right now.