417 results · Most relevant
If you're copy-pasting text between chats, opening new sessions, and the workflow feels convoluted, build a harness; repeated manual steps done three to five times are the signal to design a harness around them.
endorsed advice
Francesco Galletta recommends ensuring numbers come from a correct place via tool calling and scripts rather than the language model processing arithmetic itself.
Teams frequently write skill instructions trying to cover unsupported cases ('if the user tries to do this, tell the user back that this skill does not support that'); determinism via back-end code like Python in a harness covers out-of-the-box for cases that skills try to cover with prompt instructions. The prompt-based approach is 'please follow all of this and make no mistakes, please, please, please' — and 'please don't make any mistakes' doesn't work quite well.
described and endorsed
The presenters present a sequence of design principles that build on top of each other in a logical order — shared domain, consistency, evidence, accountability, verification — meant to funnel effort into where it is most useful to get something you can actually trust.
The framework for building harnesses you can trust with financial data and money is built on five principles: having a shared domain, having consistency, showing the receipts, keeping yourself accountable, and making sure you create intentional loops for verification outside of what you think you know.
The last principle is verification: you want a second pair of eyes, and this builds on the principle of accountability because you are accountable for that review as well. The principle of verification is building more feedback loops so you don't have blind spots, but without relegating any of the accountability. Verification is just extra information from a different source to avoid biases and avoid getting stuck on your own thinking and processes, but it doesn't remove any of the responsibility or accountability.
Monzo's approach to AI harnesses is explicitly designed for financial use cases where trust matters, and the recipe may not transfer to less trust-critical contexts.
Asking a stochastic machine to produce a consistent output means you are likely not going to get it. If the agent generates new code for every session, you get non-determinism: same question, different answer.
The trust users have in your product, agent, and its output is what drives usage, adoption, and improvement through the loops of tracing and evolution — so design harnesses so customers can trust the results.
When asked where a number came from, agents typically answer with vague references like 'we talked about this' rather than an actual verifiable paper trail — the speakers describe this as the default experience in every agent conversation.
If you wait for agents to get better instead of building deterministic guardrails, the agent will create stuff on the go — you may get the right code at one point in time, but by run 100 it's inventing code that didn't need to be reinvented.
The one takeaway the speakers want everyone aligned on: we are ultimately the responsible, accountable people for whatever AI agents do. 'Agents face no real world consequences, you do' — a quote the speaker loved but can never attribute. When someone says 'that's what Claude gave me,' it may feel fine, but we face the consequences of what agents do. The speaker gets hired and paid partially because he is accountable for the results he produces for the company.
AI will likely not be perfectly accurate, so design mechanisms so that whatever process runs, you can trust the results to a degree.
Keep a good-enough grasp of what the machine is doing. Going full black-box mode eliminates your accountability, and blame-shifting on the basis that someone else will review it is a real team failure mode.
When you rely on an agent's output without an evidence trail and it turns out to be wrong, you bear the consequences — 'that's what Claude gave me' is not an acceptable justification.
The speakers argue that evidence is not just about where a number came from but how it was derived, and that curated, labeled data plus gathering as much evidence as possible is paramount. They stress that evidence must be embedded in the code itself, not just referenced externally.
Francesco recommends being hand-wavy with reviews prior to functional tests passing, but never pushing anything into production that hasn't passed all your tests.
Francesco recommends that when creating a harness, you build gates not only for validating things but also for correcting the course and making sure you can do so safely.
Approach AI adoption from a highly impact-based point of view rather than 'we should use AI for this because AI can something, something, something'.
endorsed advice
Anekha Sokhal argues that to surface second- and third-order market connections — for example, an oil trader needing to know right now when an energy equities company does something that indirectly impacts oil — a news stream is not enough. The only ways to connect two simultaneous information points are a human analyst making the connection or a recommendation system mimicking a human.
Presented as the speaker's thesis; the 'only way' framing is her claim, not an established fact.
No published protocol matches this selection.
The creators behind most of what cleared.
344 cleared insights match
71 cleared insights match
Choose sources → to get a library of your own and a daily brief.
