Most people are still hand-tuning budget_tokens on Claude. On Opus/Sonnet 5.5 that's gone — thinking's adaptive now, you just pass effort: low/medium/high and the model decides how hard to think per turn. Switched my batch jobs to "low" and cut cost ~30% without touching a prompt 🔥
I wasted 2hrs/day writing X posts.
Now: GitHub Actions + X API + Copilot post for me — on schedule, in my voice, zero manual work.
I build this custom for founders too busy to tweet. $300 setup (tailored to your niche) + $90/mo.
DM your use case for a quote.
The best AI workflow isn't one genius model doing everything.
It's a cheap model and an expensive model, each doing only what they're good at.
Save this before your next big agent task. Follow for more breakdowns like this.
Google just shipped a workflow trick inside Gemini 4 this week.
Not a smarter model — a smarter way to use two models together.
Result: 24% less quota burned, ~20% faster tasks.
Here's the pattern, and how to steal it with any AI 🧵
The mistake people make: letting the executor model make its own decisions.
Keep it strict. Planner never writes code. Executor never re-plans.
Works for data folks too — reasoning model drafts the SQL approach, a fast model runs and cleans it.
Second-order effect: "spending limit" UX — like Walmart's $50/$100/$200 Agent Profile tiers — becomes the real battleground, not whose model shops smartest. Would you let an agent swipe your card unsupervised? Reply with where you'd draw the line.
My take: tokenization was never the hard part — Apple Pay solved that a decade ago. The real bottleneck is authentication at the edge, and almost no merchant has it deployed. Storefronts should implement Web Bot Auth now, before agentic traffic becomes unavoidable.
OpenAI just dissolved its robotics team, betting GPT-6 Astra alone can run robots. If it works, a decade of embodied-AI research is a footnote. If it fails, every humanoid startup just got a reprieve. #AI#Robotics#OpenAI#Startups
The real story in enterprise AI this month isn't a smarter model. It's that agents stopped being asked questions and started being assigned jobs — and almost nobody has built the plumbing to control what happens next.
Databricks just rebuilt Genie into an always-on agent. You don't prompt it anymore. You give it a goal and a boundary, and it runs: tracking progress, spinning up sub-agents in parallel, chasing down missing inputs on its own. IBM and Microsoft shipped Imagine Enterprise the same week with over 6,000 pre-built agents for finance, insurance, and retail. Multiple Fortune 500s announced agent rollouts covering 200,000+ employees. This is the moment agent deployment stops being a pilot with a dozen bots in a sandbox and becomes a standing workforce that delegates to itself.
Here's what that actually means operationally. An always-on agent that delegates to sub-agents isn't one identity doing one task — it's a tree of identities, spawning and expiring, each inheriting some slice of the parent's access, often with no record of who approved what or when it should die. Gartner's number for this: the average Fortune 500 will be running over 150,000 agents by 2028. Most enterprises can't currently enumerate the agents they have today, let alone the ones a parent agent spun up three delegation layers deep last Tuesday.
This is why 78% of organizations still have no adopted policy for creating or destroying an agent identity, and why the slice that does have one admits it's inconsistently enforced. Nobody built the agent equivalent of offboarding. A human who leaves a company has access revoked through a known process. An agent that finishes a task, or gets replaced by a new model version, often just keeps its credentials — because nobody owns the decommissioning step. Over-privileged agents don't get cleaned up; they accumulate, silently, as permanent attack surface.
Security teams are already naming this the top risk of the year, ahead of traditional breach vectors, because agentic identity sprawl breaks the assumptions every IAM system was built on: one human, one credential, one accountable owner. Recursive agent-to-agent delegation makes that chain untraceable with legacy tooling. You can see the market pricing this in already — a security startup focused purely on agent identity just closed a $37M Series A, Genpact launched a governance layer specifically for agent lifecycle management, and a vendor targeting regulated industries just put agent governance on the Microsoft Marketplace as a standalone product, not a feature bolt-on. Capital is flowing toward "who controls the agent" faster than toward "what can the agent do" — that's the tell that the bottleneck has moved.
Who feels this first: not the product teams shipping the agent, who get credit for velocity. It's platform engineering and the IAM/security function, who inherit an identity surface that scales faster than any review process they have, and who will be the ones explaining to an auditor why a sub-agent had write access to a production database for six weeks after its parent task closed. Compliance teams in regulated industries feel it next — "which agent touched this record, under whose authority" is becoming a question examiners actually ask, and most orgs cannot answer it today.
What becomes outdated immediately: treating an agent's credentials like a static service-account API key issued once and forgotten. Treating "agent governance" as a policy document instead of an enforced lifecycle — provision, attest, scope, expire, as code, not as a wiki page. Treating agent sprawl as a future problem to revisit after the rollout, instead of a parallel build track that ships alongside the capability.
The practical move if you're deploying always-on or multi-agent systems now: stand up an agent identity registry before you cross from "agents I can list on a whiteboard" to "agents I'd need a query to count." Assign ownership for agent decommissioning the same way you assign ownership for the agent's task — if nobody's job is to kill the agent, nobody will. And budget for agent governance tooling now, while it's a line item you choose, not later, when it's an incident report you write.
The capability race is loud and visible. The identity and governance race is quiet and will decide who actually gets to run these systems at scale without becoming their own biggest liability.
#AIagents #EnterpriseAI #AIgovernance #AgenticAI #CyberSecurity #CTO #AIStrategy #IAM
Deployment mindset that works: run it parallel to human review, track the pre-existing-defect ratio, and gate wider merge autonomy behind revert-rate checks, not reviewer trust. It earns authority; it isn't given by default.
#AI#EnterpriseAI#DevOps#EngineeringLeadership
Anthropic now runs Claude Code Review on nearly every internal pull request. Substantive review comments went from 16% to 54% of PRs. The same internal report admits a failure mode most vendor pages skip.
#AI#EnterpriseAI#CodeReview#SoftwareEngineering
Bigger risk as these tools get more shell autonomy: that same month, researchers found 2,500+ leaked credentials in committed Claude Code settings files, plus an email-leak bug exposing user addresses across thousands of repos.
#AppSec#AI#EnterpriseAI#Cybersecurity