Research

Modifier

ˈmɒdɪfʌɪə |

From linguistics — A phrase or clause that modifies meaning.

In agentic workflows — The smallest component of inference that can be:

• Composed,
• Evaluated,
• Refined, and
• Reused

I am currently investigating inference modifier generation as a context programming method to optimize context window utilization for adaptive in-context learning. To understand why — I am going to digress a bit first to explain how I got here.

In early 2025 I started getting approached by people wanting AI advice. People are naturally concerned about the future business implications of AI adoption, and there has been a lot of doom and gloom reported in the media and online. What impact will AI have on the business landscape, careers and jobs. At the time, ChatGPT prompting was falling short of its intrinsic potential, and given the people I work with are in startups, or are solo-entrepreneurs or small businesses owners, they weren’t in the position to hire large consulting firms to help them figure this out.

Question: Your into AI — what do you think?

Answer: To me, this feels like Silicon Valley back in 1997…¹

Unlike many of the critics and AI commentators, I experience existential optimism. This is my Sartre inspired — I deny your reality and substitute my own optimistic interpretation because I choose to do so as an act of faith. Yes, we are in no doubt entering a phase of creative destruction, but its already incredibly exciting and the opportunities are boundless. My biggest advice to anyone concerned about AI is — pay attention. This can’t be ignored, its impact, for better or worse is going to touch everything.

Gen AI Adoption

As a use case, the people asking me for advice seem to have a set of requirements that are different to those of larger enterprises, which often frame discussions we currently see in the media and market analysis².

It feels like there is opportunity here³, but what exactly is it? Undoubtedly the above analysis and its conclusions will age, but it is undeniable we are inside a hypergrowth tornado right now⁴.

The Divergent Use Case

As a cohort, the concerns of smaller businesses, particularly the ones I have been talking to, can be characterized as:

• Prompt leakage of trade secrets: Models ingesting and training on hard-earned domain-specific knowledge.
• Data privacy: Client data needs to remain confidential.
• Reliability: AI assistance becomes unproductive when efficiently gains are wasted because the work-product need careful fact checking.
• Skill development: Rather than a SAS wrapper, they want to build in-house capacity and know-how to roll their own solutions around their domain knowledge to protect their trade secrets.
• Upfront costs: They don’t want to become trapped by a blitz-scaled pricing model where prices rise sharply after lock-in.

My assumption here is that these small businesses want to design their own agentic workflows, but need better tools to do this — which is a fascinating user experience design challenge! It leads to an interesting set of design constraints for any software solution. Specifically by being:

1. Untethered: They want to self host models, potentially offline or in a secure private intranet.
2. Performant: They want old-school desktop application performance to avoid the SAS latency trap.
3. Discoverable: They want to leverage latent emergent behaviors specific to their problem domain.
4. Observable: They want to own their own prompts and observability to evaluate and refine them.

Being loath to burry the lead, here is a quick executive summary for those who don’t have time to read my full report.

TLDR

Product Brief: Imogen is a desktop application I am currently developing for endusers to design, evaluate and deploy reliable and data-secure self-hosted agentic workflows to productively automate everyday tasks.

Imogen is being designed to:

• Make agentic workflows secure, private and reliable;
• Meet people where they are;
• Leverage existing skills;
• Expand understanding of AI utilization;
• Inspire automation strategies; and,
• Not require years of software development or AI expertise.

Imogen places a premium on good user experience design. So it moves beyond the limitations of so called no-code visual design tools for agentic workflows. At the moment it seems like every other frontier AI company wants to replicate the success of n8n and Claud Code. Vibe coding makes imitation and feature parity cheap. Imogen is not like this. The businesses I talk to have no need for an MCP service that can tell you the weather in Tokyo. It’s just not a thing.

What they seem to want is:

1. A familiar interface that is quick to get started to test out an idea, yet able to handle scale and complexity with few limitations.
2. Agentic workflows able to capture and refine intent with context so models exhibit goal oriented behaviors that align with their business or personal objectives.
3. Design tools that map domain knowledge and operational experience into productive workflows capable of performing and completing long-running, open-ended tasks.
4. Delegated authority with checks and balances so models can make appropriate decisions on their own.

Objective: increase the ratio of AI value from human input.

I remember what it feels like to be at the frontier of a technological revolution with venture capital funding and the pressure to “sell picks and shovels during the gold rush”. Is there a first mover advantage here — maybe, maybe not, who knows? Lots of really smart people are all trying to solve the same problems. Many false signals and missteps are going to be inevitable.

What I do know is that the much hyped primacy of technologies like Netscape Navigator, Adobe Flash and Java, withered on the vine, while Linux, Apache and many other IETF Standards⁵ based open-source technologies have stood the test of time because they were transparent and useful. This is the prevailing wisdom that guides me now⁶.