Out of the box gives you ability. A built agent gives you your business, working.
A model's ability comes from the vendor. An agent built around your data, your rules and your record makes that ability yours. The research behind why, and what we build.
A model's ability comes from the vendor. An agent built around your data, your rules and your record makes that ability yours. The research behind why, and what we build.
Contents
This one started in a conversation. We are partnered with vertbuild.ai on an AI enablement project. In a working session with Nat Bell, Director of Operations at Vert Energy Group, our founder mentioned that open models can have their safeguards removed altogether. "Really, that's interesting," Nat said, and a moment later: "You can't unsee what you see." From there the two of them ended up somewhere neither had planned: how much of what a model tells you is the whole picture, and how much is the top of the list. That became years of research, traced and checked, and this video. Some say AI makes the human's part smaller. We see the opposite. Nobody prompted that exchange. The conversation is what pointed the intelligence somewhere worth going, and it is what keeps it compounding.
BLUF. Out of the box, Claude and ChatGPT bring a lot of raw general ability. Anthropic says an agent is built from four components: the model, a harness of instructions and guardrails, tools, and an environment (Anthropic, Apr 9, 2026). By that definition, a login to Claude or ChatGPT gets you all four, general ones, built to work for anyone. What it does not come with is your business. An agent built the way Ena Pragma builds one takes that same vendor model and makes its ability yours: your data and definitions, your rules, what it has learned from corrections, checks against what actually happened, and a record you keep, in a repository dedicated to your agent. On the same models, adding curated skills raised pass rates in a preprint benchmark, and adding a memory store raised scores in one Anthropic database task. For one person's one-off question, a chat window is fine.
#How the line of research runs
We trace the line from 2022 to this month. Picture a model's inner state as a point on a very large map. A "direction" is a consistent push on that map: push one way and the model leans toward saying one thing, push back and it leans away. The weights that make the map belong to the model's owner.
| Date | What happened | Source |
|---|---|---|
| May 10, 2022 | Researchers show that adding a vector to a model's hidden state can steer what it writes. It is about sentiment and sentences, not safety. | arXiv |
| May 13, 2023 | A community post steers a small model by adding activation vectors into its forward passes. | LessWrong |
| Oct 2, 2023 | Representation Engineering adds to the list of things a direction can carry: "including honesty, harmlessness, power-seeking, and more." | arXiv |
| Apr 27, 2024 | Researchers post that refusal in open chat models is mediated by a single direction, and note "a large body of prior work" on extracting a direction for a concept. | Alignment Forum |
| May 23, 2024 | Anthropic turns up one internal feature in Claude and its answers follow. The demo ran for 24 hours. | Anthropic |
| Jun 17, 2024 | The paper: refusal is mediated by "a one-dimensional subspace, across 13 popular open-source chat models up to 72B parameters in size." | arXiv |
| Jan 17, 2026 | A taxonomy paper on outside safety evaluations of frontier models says evaluators "often receive limited model access, limited information, and little time." | arXiv |
| May 7, 2026 | Anthropic publishes a tool that turns an activation into readable text and says its explanations can be wrong: "The most important limitation is that NLA explanations can be wrong." | Anthropic |
On those open chat models, the June 2024 authors found that taking the direction out stopped the refusals. That edit now goes by a name, abliteration. Models with "abliterated" in the name are now common on Hugging Face: at 6:41 PM Central on Nov 19, 2026 the model list returned 9,388 of them. By their own base-model tags, 5,194 are quantized copies, 1,866 are finetunes, merges or adapters of 785 distinct base models, and 2,328 carry no base-model tag. The list is named here, not linked.
The two Anthropic rows are the same move from the other side of the weights. In 2024 the owner of the model turned the knob. In 2026 the owner publishes a tool to read a model and says its explanations can be wrong. Whoever holds the weights holds the knob.
#What a business can own
Some of a model's behavior sits in directions inside its weights, and a business on a hosted model cannot reach them. The weights are the vendor's. Everything the model works through can be the business's.
That is the difference between what comes in the box and what we build. The box is general: instructions, tools and guardrails built to work for anyone. A built agent is specific: your data, your definitions, your rules, your record.
#What we build, and what each part gives a business
At Ena Pragma we take the vendor's model and make its ability the business's own. Each part below is something the business holds.
| Part | What it gives a business |
|---|---|
| Your data and definitions | The agent works from your records, your terms and your sources of truth. A model out of the box cannot know a rule nobody told it, such as "wait twenty-one days after the last enrollment." A built agent carries that rule into the work. |
| Your rules | What it may do on its own, and what waits for your OK. |
| What it learns from corrections | A fix is written down so the agent can apply it next time. |
| Checks against the real record | The agent reads the record to see whether the thing happened, and no claim ships on confidence alone. |
| A record you keep | What it did and what it looked at, in a form you can show. |
Does that change results? Two published tests each compared the same models with and without one addition: one with and without curated skills, the other with and without a memory store. Both show higher scores with the addition:
- A preprint benchmark, SkillsBench, first posted Feb 13, 2026 and updated Jun 14, 2026 with models including GPT-5.5 and Claude Opus 4.7, ran 87 tasks across 8 domains with the same models and harnesses with and without curated skills. The average pass rate went from 33.9% to 50.5%. SkillsBench, arXiv, preprint
- In Anthropic's own test, on one database task, Claude models with a memory store beat the same models with a stateless baseline. Opus 5.5 went from 0.655 to 0.838, and Sonnet 5.5 from 0.644 to 0.833. Anthropic, Nov 6, 2026
#Why it lives in a repository
Each business's agent is built in its own place, with its own repository, and the business's knowledge is kept there. Corrections are added to it, so the agent builds on what it has learned. Same vendor model, and more of what it does for you is yours: your knowledge, your rules, your record.
Vendors keep changing what they sell, and that is why the knowledge belongs with the business and not the model:
- In February 2026, ChatGPT retired GPT-4o. OpenAI, Feb 13, 2026
- In March 2026, ChatGPT moved existing conversations off GPT-5.1. OpenAI model release notes, Mar 11, 2026
- OpenAI's shutdown date for the Assistants API was Aug 26, 2026, announced a year earlier. OpenAI deprecations
- Between March and April 2026, Anthropic traced complaints to three changes in its own agent products, and says the API was not impacted. Anthropic, Apr 23, 2026
When the model changes, your knowledge base, your rules and your record do not have to. They are yours to carry to the next model.
#The published case: a check against the record
Access is one part. Checking against the record is another, and we have a published case. A client's migration script carried a note that a table was empty, verified. We checked it against the live database. It was not empty: it held the taxonomy the whole platform ran on. We held the migration, backed the table up, corrected it and had it re-verified. The client saw a migration a day late, with a note explaining why. Read the full case, published Jul 26, 2026.
#The record, and when it is a duty
For some uses the record is also a legal duty.
- If a business uses an AI system the EU classes as high-risk, it must keep the logs the system generates, to the extent they are under its control, for at least six months, unless other law says otherwise. EU AI Act, Article 26, text as at Jul 27, 2026
- For the uses listed in Annex III, that starts December 2, 2027. Regulation (EU) 2026/1744, Jul 8, 2026
- Colorado, from January 1, 2027, requires businesses using automated tools in consequential decisions to keep records to show compliance for at least three years after each decision. Colorado SB26-189, signed May 14, 2026
#What to do on Monday
- Pick one job that repeats and write down what done means in terms of your own records.
- Run it on a fixed set of your own tasks before and after any model change, and compare what it did, not what it said it did.
- Ask whoever builds it for the record: what it looked at, what it changed and who approved.
For an open-weight model, ask who published these exact weights and what was done to them.
#The line, end to end
Four years of one idea come to this: the weights are the vendor's, and what a business can shape is everything the model works through. That is what we build: the vendor's model, made yours, with your knowledge, your rules, your corrections and your record, in a repository dedicated to your agent. Out of the box gives you ability. A built agent gives you your business, working. That is what we practice, and what we sell.
#How we traced it
We followed this one idea from May 2022 to this month and checked each step against its source page's own words. The video and this post use only what held.
Related field notes: What our agents learn is a draft until we test it and An AI agent's notes can shape what it does.
Transcript
0:00 Strut: A business owner asks: why would I want an agent built for my business, when I can just log in to Claude or ChatGPT?
0:07 Tally: One argument answers it. It starts with what a business cannot reach inside a model.
0:13 Tally: May twenty twenty-two. Researchers show that adding a vector to a model's hidden state can steer what it writes.
0:20 Strut: About safety?
0:22 Tally: No. About sentiment and sentences. But it opens an idea: some of a model's behavior sits in directions inside its weights.
0:31 Tally: October twenty twenty-three. Representation Engineering adds more to the list of things a direction can carry: honesty, harmlessness, power-seeking.
0:42 Tally: April twenty twenty-four. Researchers post that refusal in open chat models is mediated by a single direction. In June their paper says it holds across thirteen open chat models, up to seventy-two billion parameters.
0:56 Strut: So take it out, and the model stops refusing.
0:59 Tally: On those open models, the authors found it did. That edit has a name, abliteration, coined by an independent developer. The person who spread the word says so himself.
1:11 Tally: May twenty twenty-four. Anthropic turns up one internal feature in Claude, and its answers follow. It ran for twenty-four hours as a research demo.
1:21 Strut: Same move. What differs?
1:24 Tally: Who holds the weights. Whoever holds them holds the knob.
1:28 Tally: May twenty twenty-six. Anthropic publishes a tool that turns an activation into readable text. Anthropic says the explanations from its own activation-reading tool can be wrong.
1:39 Tally: And a January paper on outside safety evaluations of frontier models says evaluators often receive limited model access, limited information, and little time.
1:49 Strut: So a business on a hosted model cannot reach any of that.
1:53 Tally: No. And it does not have to. The weights are the vendor's. Everything the model works through can be the business's.
2:01 Tally: Anthropic says an agent is built from four components: the model, a harness of instructions and guardrails, tools, and an environment. Log in to Claude or ChatGPT and you get all four, general ones, built to work for anyone.
2:17 Strut: And that is an agent.
2:18 Tally: Technically, yes. It comes with a lot of raw general ability. What it does not come with is your business.
2:28 Tally: Here is what we build at Ena Pragma. We take that same vendor model and make its ability yours: your data and your definitions, your rules, what it has learned from corrections, checks against what actually happened, and a record you keep.
2:40 Tally: A model out of the box cannot know a rule nobody told it. Wait twenty-one days after the last enrollment. A lead is not an accepted invite. Those are a business's rules. A built agent carries them into the work.
2:54 Strut: Does that change results?
2:56 Tally: On the same models, the evidence says it does. A preprint benchmark of eighty-seven tasks across eight domains, updated in June with models including GPT-5.5 and Claude Opus 4.7, ran the same models and harnesses with and without curated skills. The average pass rate went from thirty-three point nine percent to fifty point five.
3:19 Tally: And in Anthropic's own test, on one database task, Claude models with a persistent memory store beat the same models with a blank slate. Opus 5.5 went from point six five five to point eight three eight. Sonnet 5.5 went from point six four four to point eight three three.
3:37 Tally: That is why each business's agent runs in its own place, with its own repository. The business's knowledge lives there, in plain files, and corrections are added to it.
3:47 Tally: So the agent boots with what it has built up, and it builds on itself. Same vendor model. More of what it does for you is yours: your knowledge, your rules, your record.
3:58 Strut: What does that look like in the work?
4:01 Tally: Access is one part. Checking against the record is another, and we have a published case of ours. A client's migration script carried a note: that table is empty, verified.
4:12 Tally: We checked it against the live database. It was not empty. It held the taxonomy the whole platform ran on. We held the migration, backed the table up, corrected it, and had it re-verified. The client saw a migration a day late, with a note explaining why.
4:28 Tally: It is the habit we work to: no claim ships on confidence alone.
4:33 Tally: Vendors keep changing what they sell. In February, ChatGPT retired GPT-4o. In March it moved conversations off GPT-5.1. In August, OpenAI removed the Assistants API from its developer platform, which it had announced a year earlier. Between March and April, Anthropic traced complaints to three changes in its own agent products, and says the API was not impacted.
4:59 Tally: When the model changes, your knowledge base, your rules and your record do not have to. They are yours to carry to the next model.
5:07 Tally: For some uses the record is also a legal duty. If a business uses an AI system the EU classes as high-risk, it must keep the logs the system generates, to the extent they are under its control, for at least six months, unless other law says otherwise. For the uses listed in Annex III, that starts December second, twenty twenty-seven. Colorado, from January first, twenty twenty-seven, requires businesses using automated tools in consequential decisions to keep records to show compliance for at least three years after each decision.
5:39 Strut: Back to where we started. What does four years of one idea give a business?
5:44 Tally: Some of a model's behavior sits in directions inside its weights, and a business on a hosted model cannot reach them. What it can shape is everything the model works through.
5:55 Tally: That is what we build. The vendor's model, made yours: your knowledge, your rules, your corrections, your record, in a repository that builds on itself.
6:06 Tally: Out of the box gives you ability. A built agent gives you your business, working. That is what we practice, and what we sell.
#Sources
- Extracting Latent Steering Vectors from Pretrained Language Models, arXiv, May 10, 2022
- Steering GPT-2-XL by adding an activation vector, LessWrong, May 13, 2023
- Representation Engineering, arXiv, Oct 2, 2023
- Refusal in LLMs is mediated by a single direction, Alignment Forum, Apr 27, 2024
- Golden Gate Claude, Anthropic, May 23, 2024
- Refusal in Language Models Is Mediated by a Single Direction, arXiv, Jun 17, 2024
- Expanding External Access To Frontier AI Models For Dangerous Capability Evaluations, arXiv 2601.11916, Jan 17, 2026
- SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks, arXiv 2602.12670, preprint, Feb 13, 2026, updated Jun 14, 2026
- OpenAI, Retiring GPT-4o and other ChatGPT models, Feb 13, 2026
- OpenAI, Model Release Notes, entry dated Mar 11, 2026
- Anthropic, Trustworthy agents in practice, Apr 9, 2026
- Anthropic, An update on Claude Code quality reports, Apr 23, 2026
- Anthropic, Natural Language Autoencoders, May 7, 2026
- Colorado SB26-189, signed May 14, 2026, effective Jan 1, 2027
- European Union, Regulation (EU) 2026/1744, Jul 8, 2026
- EU AI Act Service Desk, Article 26, text as at Jul 27, 2026
- Ena Pragma, The migration that almost ate a taxonomy, Jul 26, 2026
- OpenAI, API deprecations, Assistants API, shutdown date Aug 26, 2026
- Anthropic, The components of agent memory, Nov 6, 2026
- Hugging Face public model list, searched for "abliterated", read Nov 19, 2026 at 6:41 PM Central (named, not linked)
Feedback
A new era.
Room for you.
Keep reading
- · 10 min
GPT-6.2 was more likely to leave out the obstacles it hit, and four more model moves to test
OpenAI's system card says GPT-6.2 more often omitted known obstacles in its reports. Four other model moves this week, each with a test and the catch on the vendor's page.
- · 8 min
EP Concept 001: Software grown under glass breaks in the real world. EP grows it in the wind.
A glass house in Arizona grew weak trees for lack of wind. EP's ninth core value, Wind-Grown, builds software the other way: tested in real use, with a kept record.
- · 9 min
The 8.5 cent employee
OpenAI says GPT-6 Luna earned 91.8% of rubric points on health admin tasks at 8.5 cents each. A five minute whiteboard video reads it against the first-party pages.