Agentic Systems Reading Room

Behavioural diagnosis

Research on agentic AI arrives faster than any one person can sensibly read it. My weekly scheduled research sweeps continually surface work on human–AI collaboration and AI adoption. Papers, reports and commentary also appear through social media, citations and ordinary browsing. The pile is not going to become smaller through exemplary personal discipline. It is designed to keep accumulating.

This creates both overwhelm and research FOMO. A new paper might contain something that matters to my work, but I often cannot know that until a later question gives it context. Reading everything immediately would consume time without producing proportionate value. Saving an unannotated link avoids the immediate decision but usually turns the paper into part of an inaccessible backlog.

The useful alternative is triage. I need enough contact with a source to understand what kind of contribution it makes, whether it is likely to matter and what would justify returning to it. That initial assessment is not a substitute for close reading. It is a decision about the appropriate depth of attention now, combined with a route back later.

The cumulative effect matters. If each source leaves behind a bounded, retrievable record, future work can start from earlier orientation rather than reconstructing it. The Reading Room turns a stream of new material into an accumulating research resource.

From barrier to design response

Observed problem	Underlying barrier	Design response
New research arrives every week from several directions	The inflow continues regardless of how much I have already read	Provide a low-cost triage route so new material can be assessed without becoming an immediate deep-reading obligation
Every new paper carries a small possibility of being important later	Its relevance may only become visible when another question, source or project appears	Preserve a bounded record of its subject, contribution and possible relevance so it can be recognised and retrieved later
Reading everything closely is impossible and unnecessary for my role	Different sources warrant different depths of technical and methodological attention	Separate initial orientation from later close reading instead of treating a paper as simply read or unread
Saved links accumulate without accumulating understanding	A bookmark preserves a location but not what attracted attention or how the source might connect to existing work	Store papers, reports and books alongside notes, concepts, glossary entries, questions and methods
Agentic-systems research uses broad and unstable labels	Apparently similar papers may describe systems with materially different capacities, environments and patterns of direction	Develop a design taxonomy that offers a shared structure for comparing configurations without assuming that one label identifies one kind of system

What I built

I built a working research notebook for triaging and gradually understanding agentic-systems research. It distinguishes external material such as papers, reports and books from the conceptual material developed while working with them: concepts, notes, glossary entries, open questions and methods.

The Reading Room supports several depths of engagement. A source can be recorded because I need to remain aware of it, annotated enough to explain why it may be useful, or revisited for closer reading when it becomes relevant to a specific question. Search and the different record types provide routes back into the material.

The system is primarily for me. Public access makes the research trail inspectable and may be useful to other people, but it does not turn the collection into a claim of completeness. The Room reflects my questions, my level of technical engagement and the decisions I have made about where deeper scrutiny is worthwhile.

Alongside it, I built the Agentic Systems Design Taxonomy, a public demonstrator of the design taxonomy that developed from my thesis. The Design Taxonomy and the Reading Room are related but not duplicates. One supports ongoing research work; the other makes a conceptual argument visible.

How it works

1. Research enters from several streams. Weekly scheduled sweeps surface work on human–AI collaboration and adoption. Social media, citations and ordinary browsing add further papers, reports and commentary.
2. I establish what the item is. AI can help identify the topic, type of contribution, central claim and likely relationship to my existing questions. This is orientation, not a declaration that the source has been fully evaluated.
3. I decide how much attention it warrants now. Some material needs immediate scrutiny. Some only needs a useful record. Some does not belong in the collection at all.
4. I preserve enough for later recognition. The record keeps the source findable and explains why it may matter, rather than storing only a title and URL.
5. Conceptual work accumulates alongside the sources. Notes, glossary entries, questions and methods preserve the distinctions and uncertainties that develop across multiple readings.
6. I return when a real question creates relevance. Search and structured record types let later work start with what I already established, while the original source remains available for the closer reading the new question requires.

The scheduled sweeps do not automatically populate the Reading Room. They create a regular incoming stream. Selection, interpretation and the decision to preserve or revisit something remain mine.

Important design choices

Design choice	What it needed to solve	What I built
Design for continuing inflow	Continuing discovery means the queue will keep replenishing itself	A triage workflow that treats recurring arrival as normal rather than waiting for the backlog to be “finished”
Separate orientation from close reading	A binary read-or-ignore model makes every potentially relevant paper feel like an immediate obligation	Records can preserve what a source appears to contribute and why it may matter without pretending that preliminary orientation equals full appraisal
Preserve the reason for saving	Titles and URLs rarely explain why a source seemed relevant several months later	Source records retain enough annotation and context to support a future decision about whether deeper reading is worthwhile
Keep research objects distinct	Sources, developing concepts and methodological questions play different roles in a research programme	The Room has separate areas for Papers, Reports and Books, Concepts, Notes, Glossary, Questions and Methods
Make a private working function publicly inspectable	Public presentation can encourage artificial completeness and polish	The Room remains visibly a working notebook shaped by my questions, while being publicly readable for anyone who finds the trail useful
Keep the live system separate from the conceptual demonstrator	A research workflow and a thesis argument should not be presented as though they perform the same job	The Reading Room supports ongoing triage; the Design Taxonomy site demonstrates the model and one possible use for organising research

The Agentic Systems Design Taxonomy

The Design Taxonomy started as a demonstration of the central idea developed through my thesis. “Agentic AI” covers systems with very different arrangements. Rather than asking only whether something is an agent, the taxonomy describes its configuration across three domains:

• System capacity: what the system can process, plan, remember, execute, revise or coordinate.
• Environmental opportunity: what information, tools, permissions, infrastructure and organisational processes are available to it.
• Activation and direction: how work starts, where objectives come from, how initiative and delegation operate, and how continuation or return of control is handled.

The Design Taxonomy site makes that model visible and explorable. The research section at the end demonstrates a further possibility: the taxonomy could help organise literature by the aspects of agentic systems that each source examines. That may eventually support the Reading Room, but it is currently a conceptual demonstration rather than an automated or comprehensive research pipeline.

I’ve written more about this in Designing agentic systems.

View the Agentic Systems Design Taxonomy ↗ (opens in a new tab)

How I worked with AI

AI makes the early stages of research triage much cheaper. It can help inspect a newly surfaced paper or report, identify its apparent focus, prepare a bounded description, suggest relevant concepts and make the resulting record searchable. Scheduled research sweeps also extend what I encounter beyond the sources I happen to see in my own feeds.

Those functions do not remove the distinction between orientation and evidence appraisal. I decide whether a source matters, what depth of reading the question requires and whether the generated account represents the source fairly. When a claim becomes consequential to my analysis, the original material must be examined rather than replaced by a fluent summary.

I also used AI to turn an evolving information model and a thesis taxonomy into two working public artefacts. Rapid implementation made it possible to test whether the categories, navigation and visual representation actually supported the research task, then revise the design while the reasoning remained fresh.

What this project demonstrates

Capabilities applied

Capability	Evidence in this project
Behavioural diagnosis	I identified that research FOMO was being sustained by a false choice between reading every potentially relevant paper closely and losing it in an uninformative backlog. The design introduces proportionate attention as a third option.
Research triage	Weekly scheduled sweeps and social-media discovery create a continuing intake. The Reading Room separates preliminary orientation, selective preservation and later close reading rather than giving every source the same treatment.
Information architecture	Papers, Reports and Books remain distinct from Concepts, Notes, Glossary, Questions and Methods, so external evidence and the structures developed around it can accumulate together without becoming interchangeable.
Conceptual modelling	The Design Taxonomy site translates the thesis argument into a three-domain taxonomy covering system capacity, environmental opportunity, and activation and direction, with more detailed dimensions beneath each.
Iterative AI prototyping	I used AI to turn both a research workflow and an abstract taxonomy into public, testable artefacts. The Design Taxonomy site’s research section then tests a secondary use for the taxonomy without claiming that it is already the Reading Room’s operating structure.

Capabilities extended

The project is extending my ability to design proportionate research workflows: deciding what needs awareness, what needs preservation and what needs genuine close reading. It is also developing the relationship between evidence management and conceptual modelling. A taxonomy can provide useful orientation across a literature, but only if it remains open to papers that complicate or expose gaps in the model.