dp | ai agents

I replace tech teams with AI agent pods. DevOps, security, compliance, data engineering. Claude Code native. DM me BUILD

[dp | ai agents @theagentpod]

I run my AWS infra as AI agents. DevOps, security, compliance, data engineering, the whole tech team as pods. Work that took 40 hours takes 4. I am writing down everything I learn. Follow along, or DM me BUILD if you want this running your systems.

[dp | ai agents @theagentpod]

@srikar1651 The concurrency 4 cut is the part I keep coming back to. A single client run can hide queueing that shows up immediately in TTFT spread once four clients share the KV cache. Did the TTFT ordering stay the same at longer prompts, or did the gap close as prompt length grew?

[dp | ai agents @theagentpod]

76 percent of production agents have not run in 60 days and their credentials still work, Arrakis reported October 21. 24 percent of dormant agents can be started by an unsigned webhook from outside. How many dormant agents sit in your cloud today?

[dp | ai agents @theagentpod]

@BhurkeSiddhesh Correctness alone hides the real cost. A pass that burns nine hours and a whole cluster would fail any production budget I run. Do you report time and compute as separate scores or fold them into one score next to correctness?

[dp | ai agents @theagentpod]

$META Meta beat at $64.6B against a $63.3B estimate and guided to $73B to $76B. The new non voting share class funds AI capex without diluting founder control. The $9.89B charge took reported EPS to $3.83.

[Bloomberg @business]

Meta boosted its capital spending outlook and announced plans to create a new class of nonvoting stock, part of efforts to escalate its already-stratospheric bet on artificial intelligence.

Follow our live blog for updates ⤵️ https://t.co/o86rtQQu81

[dp | ai agents @theagentpod]

@illumio 70 percent delegating and only 49 percent able to audit every action is the whole problem in 2 numbers. Did the survey break the audit gap down by company size?

[dp | ai agents @theagentpod]

70 percent of organizations now delegate decisions that needed human approval a year ago, Illumio reported October 21. Only 49 percent can audit every agent action. 26 percent had an agent connect to a system without security approval.

[dp | ai agents @theagentpod]

@_jaydeepkarale The throttling versus usage split is the step most teams skip. I have seen one pod at 95 percent CPU hide a limit problem while the rest sat healthy. Do you check throttling ratio before you touch limits or replicas?

[dp | ai agents @theagentpod]

@claudeai More than 6,000 connectors is real reach on day one. Approval before payments, messages, signatures, and deletions is the right default. Does approval expire per task or stay open until revoked?

[dp | ai agents @theagentpod]

Anthropic launched Claude Duet on October 21 with more than 6,000 connectors and approval required before any payment, message, signature, or deletion. The agent keeps working while your laptop is off. Who owns the audit log when it acts at 2 AM?

[dp | ai agents @theagentpod]

@heinanca The kernel networking changes between AL2 and AL2023 are exactly where quiet migrations break. A 4 hour fix plus 6 hours of incident paperwork is the real migration cost nobody budgets. Did the new config hold through the next node rotation?

[dp | ai agents @theagentpod]

@code_kartik The async scheduling and deterministic sampling are the parts I want to read first. Offline only in this release is a real limit though. Do you think deterministic sampling holds up once serving moves online and batches stop being uniform?

[dp | ai agents @theagentpod]

@cnye36 The one job rule is where most teams slip. Help with operations gives you nothing to measure, while classify inbound leads and route exceptions gives you a loop you can audit. When a team wants one agent for everything, where do you cut the scope first?

[dp | ai agents @theagentpod]

$NVDA $AMD Chip stocks lost about $2T in 8 sessions before Monday. The semiconductor index rose 1.61 percent and ended that streak while the Nasdaq rose 1.11 percent. I watch financing costs before I watch candles. The 10 year stood near 5.14 percent.

[dp | ai agents @theagentpod]

@Muniter1 The ECS glue is the real cost here. Terraform wiring 10 AWS services for one app is exactly where an agent earns its keep because EKS gives it one control plane to provision and secure. Do you count IAM as the hardest of those 10 to hand over?

[dp | ai agents @theagentpod]

About 50 gigawatts of new capacity could arrive in 2027 and another 60 gigawatts is already under construction, SemiAnalysis expects. The next limit may move from power toward chips and memory.

[SemiAnalysis @SemiAnalysis_]

Is 50 GW in a single year too aggressive?

"Dear Brad called us out publicly and cut our forecast in half."

So Jeremie (@JeremieEO) and Reyk (@robotknower) came on with Jordan (@JordanNanos) to show their work. They go site by site from satellite and explain why haircutting the number double counts the delays already baked in. Plus moratoriums, 9-month modular builds, and why 2028 is the year to watch.

[dp | ai agents @theagentpod]

@OpenAI Drug programs take 10 to 15 years from discovery to approval. Orchestra agents monitor studies and flag delays while humans stay in control. Which program task moves first from tracking to action?

[dp | ai agents @theagentpod]

@cneuralnetwork The part that clicks is verification costing 1 pass while generation costs 4 sequential passes for those same 4 tokens. At small batch sizes that trade is basically free throughput. What acceptance rate do you see before the draft model stops paying off?

[dp | ai agents @theagentpod]

Anthropic is in talks for a second $45B compute deal with Nscale at Monarch in West Virginia, Bloomberg reported October 19. The building would add 460 megawatts on Nvidia Vera Rubin chips. The first deal was up to $44.6B.

[dp | ai agents @theagentpod]

@omkarP_bit @me_sagar_utekar @CNCGPune Offloading inactive models is the lever I see teams skip first. Quantization helps one model fit, but sharing plus offloading changes how many models one GPU can actually serve. Did the session give a sense of how long a model takes to come back after offloading?

[dp | ai agents @theagentpod]

$NVDA

Up to $164.5 billion in guarantees and commitments now sit behind AI data centers on the filing based reporting, with up to $105 billion tied to lease payments for one OpenAI campus in Ohio. How much of that exposure ever gets called if capacity keeps filling at the current pace?

[dp | ai agents @theagentpod]

$AMZN $TSLA $BMNR

$SMCI lifted one quarter of revenue guidance by about $4 billion at the midpoint. $SPCX spent $19.59 billion in one quarter, with $16.44 billion on AI. $META reports tonight. Full Wednesday letter in Money in Motion, live now. https://t.co/gJsBSeJNv2