Enhans

AI for Real Work
Enterprise AI agents that run the work end to end.

[Enhans @enhansai]

https://t.co/063LkiXC5j

[Enhans @enhansai]

[Think with Enhans: July 9, 2026]

The Claude Code team is framing agent work as loops: turn-based, goal-based, time-based, and proactive.

The signal is clear. Reliable agents need stop conditions, feedback cycles, and workflow boundaries designed before execution starts. Prompt quality matters. Operating structure is what makes the work repeatable.

#AgenticAI #EnterpriseAI #AgentOS

[Enhans @enhansai]

@cyrilXBT The AI Engineer this article describes is exactly what we call an FDE at Enhans. Not model training. Building AI that actually runs inside enterprise workflows. One layer this guide skips: ontology. We covered this in our last piece on FDEs.

Full article in our profile.

[Enhans @enhansai]

[Think with Enhans: July 7, 2026]

Fable cut Claude Code costs by 59 to 70% with a strange move: convert large code context into PNG images, then let the model read it with OCR.

The bigger signal: agent cost is now an architecture problem. Teams need to decide what stays as text, what gets compressed, and what context is recent enough to preserve. At scale, token efficiency becomes part of system design.

#AI #AgenticAI #EnterpriseAI

[Enhans @enhansai]

[Think with Enhans: July 3, 2026]

AI generates the output. Context determines whether it is useful.

Atlassian embedded brand intent and design logic directly into prompts so AI-generated UI could carry actual meaning. Without that structure, the output looks plausible but means nothing.

The same is true for every enterprise AI use case. The context layer comes first. The model comes after.

#EnterpriseAI #AgenticAI #Ontology

[Enhans @enhansai]

@JamesOClark Mostly right. Missing one step before all of it.
Before you navigate the trench, you have to define what you're actually solving in it. Legacy systems and messy data are navigable. Building the wrong thing after you've navigated them is not.

[Enhans @enhansai]

Agreed. And even the right engineer fails without the right first move: define the problem before building anything.
The skill most FDE transfers lack isn't technical. It's resisting the instinct to start building before the problem is correctly scoped. That part doesn't come from reassignment.

[Enhans @enhansai]

@ajwade Consultant, builder, teacher. The order matters more than the mix.
Most FDE failures happen when building starts before the problem is correctly defined. The diagnosis step isn't in the job description. That's exactly where the leverage sits.

[Enhans @enhansai]

@stretchcloud Right signal, missing variable: what does the embedded engineer do first? Agents compress implementation. They don't compress problem definition. The FDE value is highest before the build begins.

[Enhans @enhansai]

https://t.co/SZjEPkXEJD

[Enhans @enhansai]

[Think with Enhans: June 22, 2026]

Enterprise security standards don't bend for new tools. MCP had to catch up.

Zero-Touch OAuth brings centralized authorization, single sign-on, and audit trails to MCP servers. The protocol now meets the bar enterprises already require.

The gap between promising AI infrastructure and production-ready AI infrastructure just got smaller.

#EnterpriseAI #AgenticAI #MCP

[Enhans @enhansai]

Every model will be replaced by a better one. The companies that stay ahead are the ones whose domain knowledge, judgment, and workflows are encoded in a structure they control. That structure is what makes AI theirs.
That's what we build at Enhans, through ontology.
#EnterpriseAI #AgenticAI

[Enhans @enhansai]

[Think with Enhans: June 16, 2026]

Satya Nadella proposes a test: can your company swap out a generalist model without losing its "company veteran" expertise? If the answer is no, you don't own your AI. You're renting compute on top of someone else's knowledge.

[Satya Nadella @satyanadella]

https://t.co/vLmiBKTtX3

[Enhans @enhansai]

@AndrewBolis Useful split, especially calling out Non-Agentic as just prompts. Too many teams jump straight to agents when a simple prompt workflow is enough

[Enhans @enhansai]

@__ghostfail the “local language ontology” phrase is dead on. every long context window starts inventing dialect, and a living margin glossary would help a lot

[Enhans @enhansai]

@hasantoxr 5.2% likely malicious is the line that should make every agent marketplace nervous.

[Enhans @enhansai]

@SanUvacha I’ve seen large SI teams standardize coding workflows around Claude for repo understanding and refactors, so losing something like “Fable 5” would hit delivery playbooks, not just side experiments.

[Enhans @enhansai]

@nvidia @ArtificialAnlys “dozens to hundreds of AI model calls” is exactly why this needs its own benchmark. single-call evals miss where agent systems actually get expensive and flaky.

[Enhans @enhansai]

@0xMorlex Yep, “hand it days of work and walk away” is the real shift. The Stripe example on a 50M-line Ruby codebase is the kind of scope where staged planning and self-checking actually matter.

[Enhans @enhansai]

@BarabiloT 100%. Once an agent can draft emails and route prompts, visibility into each step stops being nice-to-have and becomes the first thing security will ask for.