Luis Valencia

Microsoft MVP 10 years, Book Author

[Luis Valencia @levalencia]

@precisox Hermes

[Luis Valencia @levalencia]

@JulianGoldieSEO Over engineered

[Luis Valencia @levalencia]

@skynews357 I have an m4 pro and it doesn’t Even run gemma 4, this post is hype, show me
Benchmarks of the model you run locally

[Luis Valencia @levalencia]

@lorden_eth I prefer Tolaria

[Luis Valencia @levalencia]

The stack needs task queues, permissions, logs, evals, budgets, and rollback. Otherwise we are just giving a confident system more places to be wrong.

[Luis Valencia @levalencia]

Who can assign work?
What can the agent access?
Can it open PRs?
Can it trigger CI?
Can it patch failed builds?
Where does review happen?

[Luis Valencia @levalencia]

If another system can create and track agent tasks, the agent becomes part of the SDLC. Not just a developer sidekick. A workflow participant.

[Luis Valencia @levalencia]

Coding agents are starting to look less like chatbots and more like infrastructure.
GitHub’s latest Copilot updates point in that direction: larger context, configurable reasoning, CI failure help, and an Agent tasks REST API.

[Luis Valencia @levalencia]

Anthropic’s autonomous vulnerability discovery/remediation harness is interesting as a workflow blueprint, not a ready-made production tool.
The useful pattern:
decompose the security loop before automating it.
Threat model → scan → triage → report → patch → review.

[Luis Valencia @levalencia]

Bigger context windows are useful.
But for coding agents, I’m more interested in:
task APIs
permissions
CI feedback
logs
evals
rollback
review gates
That is where agents become systems instead of demos.
https://t.co/Y5yzZokAet

[Luis Valencia @levalencia]

Coding agents are moving from chat boxes to work queues.
GitHub’s Copilot Agent tasks REST API is the important bit here.
Once an agent has an API, the question changes from “can it write code?” to “how do we govern semi-autonomous work inside the SDLC?”
https://t.co/czkwdCAR3L

[Luis Valencia @levalencia]

@AlexFinn btw,. shouldnt Hermes Desktop automatically detect all profiles? I have around 10 different profiles and I cant see them on Hermes Desktop! it says there are no sub agents, but I wonder if I missed something

[Luis Valencia @levalencia]

@AlexFinn bye bye Telegram? not really, everytime I am walking I am using telegram with audio messages to do coding tasks while I get to my desktop! but hermes desktop is nice!

[Luis Valencia @levalencia]

Open models are not just model releases.
They are runtime stories.
Gemma 4 12B may be interesting, but real adoption will depend on the boring path: Ollama, llama.cpp, vLLM, quantization, crashes, fixes, and serving stability.
Model release != production readiness.

[Luis Valencia @levalencia]

“Capability is not permission” is becoming the core rule for agents.
A model may be able to browse, code, call tools, edit files, and trigger workflows.
That does not mean it should get broad credentials and vibes-based supervision.
Agent safety is becoming systems engineering.

[Luis Valencia @levalencia]

OpenAI deprecating Agent Builder, the Evals platform, and reusable prompts is a good reminder:
Do not build production agent workflows as if vendor tooling will stay still.
Keep evals portable.
Keep traces exportable.
Keep migration paths boring and explicit.

[Luis Valencia @levalencia]

The agent question is changing.
Old question:
Can the model solve the task?
New question:
Where does it run, what can it access, what does it cost, what gets logged, and can we roll it back?

[Luis Valencia @levalencia]

Coding agents are starting to look less like “AI autocomplete” and more like infrastructure.
The interesting GitHub Copilot updates are not only model support.
They are sandboxes, automations, SDKs, billing, and budget controls.
That is the enterprise agent stack showing up in pu

[Luis Valencia @levalencia]

@FicoGutierrez le falta tanta coherencia? A ver como porque serían las últimas? Y porque desactivar los comentarios? No tiene la inteligencia para sostener sus estupideces?

[Fico Gutiérrez @FicoGutierrez]

🗳️🇨🇴Desde Medellín, reportamos votación masiva a estas horas. Hay largas filas.

Invito a la gente a salir temprano, no lo dejen para lo último.

Tenemos una gran responsabilidad: cuidar la democracia.

Depende de nosotros que estas no sean las últimas elecciones en Colombia 🇨🇴.

[Luis Valencia @levalencia]

@djdextune This is exactly the kind of DJ workflow fix that deserves documentation. Lost playlists are brutal because they’re not just “files” — they’re memory, prep time, and gig context.

One thing I’d add: after recovery, make a quick external export/backup habit for playlists