News 4 u.

A compensation model built around output

Tenex began by paying engineers for story points rather than hours, reflecting the guests’ view that AI increases the amount of work a capable engineer can deliver. Lieberman says the model has produced wide earnings differences, though roughly fourfold rather than the hoped-for tenfold spread. He says gaming and poor-quality code have not been the dominant problems because AI strategists are tied to customer retention and growth. The unresolved issue is valuing architecture, discovery and client work that do not fit neatly into tickets.

Context becomes part of engineering work

Hess describes the MetaHarness as a shared collection of skills, project documents and command-line tooling around the codebase. Plans, acceptance criteria, decisions, visualizations and client readouts are checked into Git. The aim is to preserve not only what the team built but why it made particular choices, giving agents a durable source of context instead of relying on conversations or individual memory. The guests acknowledge that such structure may be excessive for a small project.

Reported productivity gains require calibration

Hess estimates that engineer output has roughly doubled over the period when models became better at long-running tasks. The guests also note that story-point values may need regular recalibration as models improve, creating a tension between rewarding higher output and measuring the underlying value accurately. Their productivity figures are internal estimates from the interview, not an independently audited study.

Token costs and model choice

The guests say Tenex’s token spending is roughly comparable to or somewhat above employee spending, while falling model prices have reduced the danger of uncontrolled usage. They describe using different models and coding environments for different tasks, with Opus 5.5 presented as a current default and other tools used for computer interaction or review. These product preferences and cost estimates are conversational reports rather than verified market data.

Enterprise adoption drops at deployment

Lieberman estimates that most enterprises in Tenex’s pipeline have supplied employees with a general-purpose model, fewer have provided meaningful training, and about half have deployed coding agents effectively. He puts production agents doing non-engineering work at roughly 5 percent. The guests attribute the gap to weak processes, permissioning, governance and the difficulty of turning inconsistent human work into something that can be evaluated reliably.

Automation starts with process definition

Tenex’s advice is that an organization cannot automate a process that does not yet exist in a consistent form. When employees perform similar functions differently, an agent lacks a stable target and evaluators cannot tell whether its work is correct. The proposed sequence is to standardize the workflow, define evaluation criteria and then move repeatable work into the background, with humans reviewing and correcting the results.

Vertical AI faces a software-services tension

Lieberman is generally skeptical of vertical products whose central feature is an agent performing a specialized task, arguing that one-of-one enterprise expectations can undermine software margins and expose products to frontier-model improvement. The counterargument is that specialized data, interfaces, integrations, compliance and entrenched user habits can create durable value. The Bloomberg comparison illustrates how workflow and network effects may matter as much as access to underlying data.

Build versus buy depends on the size of the prize

The Kill My SaaS example shows why internal development can appeal to smaller or more nimble organizations: the team expects a better workflow and faster iteration, not only a lower bill. The guests argue that replacing core enterprise systems is much riskier because errors can affect enormous sums and because maintenance is part of the cost of ownership. The practical test is whether the improvement and savings justify the ongoing cognitive load.

Work may be reallocated, but displacement remains real

Lieberman and Hess argue that AI can move subject-matter experts toward strategic work as routine tasks are automated, while acknowledging that some roles will disappear and that this can be devastating for affected workers. They describe reskilling as a responsibility shared by employers and employees, but also concede that workers may lack the flexibility to prepare for changing demand. Their conclusion is optimistic about aggregate productivity but not about every individual outcome.

New York’s AI ecosystem and the conclusion

The guests portray New York as a smaller AI bubble than San Francisco but a large commercial market where enterprise adoption could have broad economic effects. They see Tenex as trying to expand that local ecosystem while connecting frontier technical communities with customers. The episode closes with Tenex’s hiring plans and its ambition to make internal agent infrastructure useful in enterprise environments, especially where permissions and operational messiness remain unresolved.

Context & caveats

• The transcript is automatically generated, contains wording errors, and does not consistently identify speakers; attribution follows the surrounding conversation and may be uncertain.
• Product names, model preferences, customer adoption rates, productivity estimates and cost figures are claims or estimates made during the interview, not independently verified measurements.
• The original episode contains promotional discussion of Tenex and its hiring; those elements are included only where they clarify the company’s stated direction.