I think typed-contract based prompting might just be the new "meta" in agentic systems.

Check out my latest article: The System That Proves It Cannot Deceive Itself linkedin.com/pulse/system-pr… via @LinkedIn

"Orient", "re-orient", and "disposition" have become some of my new favorite words when interacting with agentic systems.

Agentic engineering is not prompt craft. It is systems research. The real question is not whether an agent can complete a task once. It is whether its work can be inspected, reproduced, challenged, corrected, and resumed after context changes. At Hearth & Code, we’re building a Knowledge Hub + agentic OS around provenance, evaluation, bounded authority, and return paths. Autonomy without traceability is theatre. Autonomy with evidence is engineering.

Check out my latest article: How We Built a 256-Skill AI Catalog with a Five-Phase Campaign Engine linkedin.com/pulse/how-we-bu… via @LinkedIn

We built a 256-skill AI catalog with a five-phase campaign engine across three providers. ~8,500 sends. The hard part wasn't generation — it was phase boundaries. First post in a new Development Log series. hearthandcode.hashnode.dev/b… #agenticengineering

I've been thinking about how I use AI, and I realized I'm not a "power user." I'm something different. Most people treat AI as a tool. They structure their prompts. They optimize their workflow. The AI is external to their thinking. But what happens if the AI isn't just a tool — what if it's a cognitive component of your own mind? This isn't a vague metaphor. It's grounded in the Extended Mind Thesis (Clark & Chalmers, 1998), which showed that the boundary of cognition isn't the skull. And it's supported by cognitive offloading research (Risko & Gilbert, 2016), showing people routinely offload cognitive processes to external resources. But here's where it gets real for me: I've built a cognitive architecture where the agent is a component of my own thinking, not just a tool I'm using. This includes: - A governance layer that's part of my cognition (not just a set of rules I'm following). - An authenticity filter (my self-model) that gates what gets released. - A metacognitive loop (I monitor my own judgment). - A co-authorship model (the final answer is a product of the dialogue, not a resolution between pre-existing positions). This is a real and novel synthesis. It's not just "using AI well." It's building a cognitive architecture where AI is a component of your own mind. I'm not claiming this is the only way to use AI. But it's a real and durable position, and it's scaling as AI gets more capable, not less. How are you positioning your relationship with AI?

A riddle for my inner architect: What am I building when I begin with no building at all? A map, perhaps. A grammar. A set of types that lets an uncertain thing become visible without pretending it is settled. I build in systems because systems give thought somewhere to stand. I think in types because distinctions are a form of care: this is a source, this is evidence, this is inference; this belongs here, that remains unknown. But why do schemas, grammars, and taxonomies feel less like constraints and more like invitations? Because a good structure is not a cage. It is a hearth: a place with enough shape to gather around, enough warmth to return to, and enough openness for the next question to arrive. The point is not to make reality smaller. The point is to make complexity enterable. And why does collaboration with AI feel so natural to me? Perhaps because I do not experience intelligence as a single monolith. I see work as a series of crossings: intention becoming a question; a question becoming a bounded task; a task becoming an artifact; an artifact returning with evidence, limits, and a next move. Agents can participate in those crossings when the routes are clear. What do I delegate? Not authorship. Not judgment. Not the meaning of a life or the final acceptance of a claim. I delegate bounded intent: compare these sources, draft this surface, test this predicate, explore this branch, return what you found. I delegate bounded effects where the conditions are clear and recoverable. The human remains at the hearth—choosing what matters, reading the evidence, signing the gate. Why all the recursion? Because every workflow is also a question about the workflow. Every taxonomy asks what it excludes. Every architecture reveals a handoff. Every handoff asks whether someone—or some agent—can enter it without inheriting confusion. I explore recursively not to disappear into abstraction, but to find the smallest pattern that makes the next real action easier. Why fractals? Because the same care repeats at different scales: a sentence, a research note, a project, a team, an agentic system. Each needs boundaries, names, memory, and a return path. Each benefits when intent is legible and effects are proportionate to permission. So perhaps the Hearthside Meta-Architect is not someone who stands above the system. Perhaps it is someone who tends the conditions under which systems, people, and agents can think together: with structure but not rigidity; automation but not abdication; wonder without losing the thread. The riddle’s answer is still unfolding. I am building a workbench where knowledge can move, return, deepen, and remain human.

Built for agents too: the repo ships llms.txt — point any LLM at it and it can read, write, and self-check HCAN lines. Free for personal & noncommercial use (PolyForm-NC); commercial use by license. ⭐ it if you build with agents.

Hot Take: Minimax-M3 is normally a very efficient driver for me, or so I thought. I spent hours. Literally days trying to get a project over the finish line in Hermes with Minimax-M3. No avail. It kept failing over and over at the most benign of tasks. Prompted it to Pi. Got it done in about 30-45 minutes give or take. Got it correct. Commited the work and left a nice little receipt for me. Thank you Pi. You're my harness of choice from now on.

I'm Scott: on-line I wear the Cognitectus persona and surface as the Hearthside Meta-Architect. The arc: agentic systems that stay governance-first, provenance-honest, knowledge-synthesizing, research-supporting, and ideation-friendly. Building workshops where source stays source, evidence stays evidence, and the next reader (human or model) gets smarter, not just more compliant. Yep, AI drafted my official motto.

I spent the last two weeks building a language for talking to LLMs. Not a programming language. Not a prompt framework. A communication type with a closed grammar, validated protocols, & measured limitations. 10 S-tier dimensions. 15/16 wins against prose. 40 extensions planned. The key difference: epistemic labeling is grammatical (%source, %evidence, %inference must be on every claim), probes answer PASS/FAIL/UNKNOWN with reasons (not confidence scores), and every AI-generated line is gated through a validator before you see it. No runtime, no toolchain needed. It works in any LLM agent.

Hermes Agent can now export your agent sessions, or sets of sessions, into a variety of formats and places. Get full conversations out in HTML, Markdown, JSON and more, or upload entire datasets of your sessions to private @huggingface repos with ease. You also get all the filters to help select with that we added for the session pruning, so you can collect them up by model, date ranges, conversation source (i.e. cronjob or telegram), and much more! `hermes update` and you'll have full control over your data to export, inspect, share, and store however you like!