The durable AI stack is moving beyond the model layer
Today's signal is that the model is becoming one replaceable component in a larger operating system. The compounding advantages are moving into routing, execution environments, evaluations, credentials, and cost controls—the layers that determine whether intelligence can be deployed reliably at enterprise scale.
IBM argues that model routing is a cluster-level optimization problem
IBM Research says routing economics cannot be reduced to choosing the cheapest capable model. Cache state, concurrent workloads, serving infrastructure, and the way requests interact across a cluster can dominate the marginal price of an individual inference call.
Evaluate routing against the full serving system. Track cache hit rates, queueing, accelerator utilization, latency, and workload interference alongside model quality and token price, then optimize for cost per successful task rather than cost per request.
Source: Model Routing Is Simple—Until It Isn’t — IBM Research
An autoresearch agent reportedly beat a harness tuned by humans for two years
Zhengyao Jiang describes running an autoresearch agent on its own harness for eight days. The resulting system reportedly outperformed the human-tuned version on a held-out benchmark, reduced prompt size by 16x, introduced a new search method, and added layered defenses against reward hacking.
Self-improving systems make the evaluation boundary more important than the optimization loop. Keep held-out tests isolated, version every harness change, constrain what the optimizer can modify, and require independent review before a self-generated improvement reaches production.
Source: Zhengyao Jiang on Thread Reader
Runta is rebuilding CPU execution around the needs of autonomous agents
a16z led Runta's seed round around the thesis that agents need a purpose-built computer rather than another human-oriented container. Runta is developing an execution layer for local and cloud workloads with sandboxing, policy controls, file-access restrictions, spending limits, and efficiency monitoring.
Agent infrastructure should expose policy as a first-class control plane. Require per-task resource budgets, explicit filesystem and network permissions, isolation boundaries, and complete execution logs before granting an agent a general-purpose runtime.
Source: Investing in Runta — a16z
1Password keeps credentials out of the model while letting Claude act
1Password introduced credential delegation for Claude agents. The design lets an agent sign in and complete authenticated tasks while passwords and one-time codes remain outside the model, its memory, and Anthropic's systems.
Give agents capabilities, not secrets. Use a broker that releases narrowly scoped, short-lived authorization at execution time; keep raw credentials out of prompts and logs; and make every delegated action revocable and auditable.
Source: 1Password on X
Ramp disclosed a $1.5 million week of AI token spend
Ramp CEO Eric Glyman says the company's AI token bill grew from a rounding error to more than 10% of payroll within a year and reached $1.5 million in one week in May. He also highlighted how difficult it was for finance teams to identify which teams and workflows were driving the cost.
Treat model usage like cloud infrastructure spend. Attribute every request to a team, product, environment, and workflow; set budgets and anomaly alerts at those levels; and connect spend to task outcomes so high usage can be separated from high waste.
Source: Eric Glyman on X
Moonshot announced Kimi K3 as open frontier intelligence
Moonshot AI announced Kimi K3, a 2.8-trillion-parameter multimodal model with a one-million-token context window. The company says its Kimi Delta Attention enables up to 6.3x faster decoding at long context and that open weights will arrive on July 27.
Open weights can reduce dependency on any single model vendor, but the durable asset is the layer above them. Keep evaluations, routing, proprietary context, and application interfaces model-agnostic so a new model can be tested and adopted without rebuilding the operating stack.
Source: Moonshot AI on X
Bottom line
The model layer is becoming more capable, open, and interchangeable. That makes the surrounding system more valuable: routing determines real inference economics, agent runtimes enforce boundaries, credential brokers control authority, evaluation harnesses govern self-improvement, and spend telemetry keeps autonomy financially legible. Enterprise advantage will compound in those control layers.
Sources
- Model Routing Is Simple—Until It Isn’t — IBM Research
- Zhengyao Jiang on Thread Reader
- Investing in Runta — a16z
- 1Password on X
- Eric Glyman on X
- Moonshot AI on X
Chompute Daily is a weekday briefing on the operating consequences of enterprise AI news. Reporting belongs to the linked publishers; analysis and operator takeaways are Chompute's.