AI engineer and builder Swyx
For users constrained by throughput, Swyx suggests trying low reasoning effort with a critique loop rather than simply turning up the effort level. He points to swyxio/skills/babysit-runs for the approach. He then notes that Karpathy uses medium effort on Opus 5.5 for Hackertube, which launched that day, as an example of different choices. He compares strong opinions about effort levels to the tabs-versus-spaces debate and favors tuning to the task and personal experience. These are his suggestions and observations about another project, without a controlled comparison establishing that lower effort is better for every task.
Box CEO Aaron Levie
Aaron Levie proposes a test for AI use: if you could have produced the same result before AI, you may not be pushing its capabilities far enough. He argues that the goal should move beyond doing existing work a little faster toward qualitatively different outcomes. For prototypes, he wants not just working, interactive software but features that would previously have been difficult to deliver; for architecture, performance improvements that otherwise would not have been attempted. He sets ambitions for research across more dimensions and business processes that operate differently and learn from past failures. These are goals, not guarantees that a tool has covered every fact or will never repeat a mistake. He also acknowledges uneven intelligence and argues for combining AI’s strengths with human strengths.
Latent Space — Using Agents for Long-Horizon Financial Work — Aman Singh, Arca
Key takeaway: Advancing work across several days requires an updated picture of the client’s situation, with explicit human and system boundaries for money movement, external communication and changes to operating rules.
Arca AI platform lead Aman Singh explains that the company provides wealth management services directly rather than selling software to other advisors. AI handles analysis and work around the advisor, while clients primarily interact with a person. He describes the model as “agents prompt people more than people prompt agents”: the system proactively tracks deadlines, changes in a client’s situation and who needs to participate next.
Handling a tax bill may span ten days. An agent can analyze asset allocation and propose how to raise the required amount, but it cannot autonomously move money, liquidate assets or contact clients directly. It brings in an authorized person with a plan they can review and adjust. Access to a custodian’s API does not imply giving a model authority to execute trades; Singh explicitly says the firm is not doing autonomous trading.
Memory systems connect the long-running work: what is being serviced, who owns it, what has happened, and which actions matter in the near and longer term. Replies in an existing email thread go to the same agent, while an intervening phone call may wake a new agent that reads the shared state. A daily reflection process updates this picture so new signals can contribute to existing work, rather than merely restoring an old chat.
Controls sit at capability boundaries. Singh describes deterministic execution restrictions, checks on database writes, injection checks on incoming material and sensitive-information checks before data leaves the system. A sandbox handles computation; code-based queries are translated into a constrained format before reaching the backend. Client-facing artifacts are also checked for account information and promises of returns. These are architectural claims, not independent verification of security or compliance.
Model choices follow internal task benchmarks and complexity tiers, without requiring advisors to choose a model. Financial facts can go stale, so verifiers nudge agents toward a curated knowledge base or a fresh search. Staff explain edits and cancellations; feedback is paired with the original trace, and a nightly meta agent suggests changes to tools, knowledge or prompts. Engineers still review prompt changes because a model may turn each local fix into an expanding rule. Experts can also maintain knowledge directly in the product, creating a traceable improvement loop rather than automatically shipping every proposed change.