This is one of my favorite scenes in The West Wing, with one of my absolute favorite characters (Allison Janney as CJ Cregg is simply marvelous).
CJ works for the President and, as part of her job, takes a meeting with the Organization of Cartographers for Social Equality which wants the president to enact legislation so schools stop teaching the Mercator Projection of maps, and instead move to the Peters Projection. The President, of course, has more serious things to attend to; his attention can only be directed to so many things at once.
This problem is not new and pops up in many places. Whether you are a president triaging intel and advice, a VC sifting through thousands of pitch decks, a researcher deciding what ideas to build on or collaborators to work with, or a popular person debating which friend to get a drink with tonight, allocation of attention is fundamentally a decision-making-under-uncertainty problem.
Will this turn into a fruitful collaboration? Will investing here make me lots of money? How will taking this meeting inform our administration's policy agenda? Will I have fun? Will this friendship last, or are these folks going to leave NYC?
The President solves this through trusted advisors to whom they can delegate their attention. Venture capitalists solve this by avoiding a flood of front door email pitches, instead preferring introductions from other investors or founders they've already built trust with. Science has similar mechanisms, with researchers relying on colleagues, collaborators, conferences, journals, and informal recommendations to decide what deserves consideration. After all, failed research directions can be costly to not just their time but their careers.
Under abundance, this issue of deciding which possibilities deserve our attention, understanding why they matter, and turning them into action becomes considerably harder. The systems that do this inevitably create walls that determine what is and is not considered. Though they may seem exclusionary, this quality fulfills a natural role in making time-bounded decision-making more tractable: trusted networks couple selective attention with the context, credibility, and pathways to authority needed for action.
In the last essay, I treated finding and understanding as the final gates through which a finding becomes buildable. Here, I want to explore what happens afterward.
Unlike classic discovery systems which surface ideas or people that might be relevant, emerging systems around actionable possibilities go deeper and look for arrangements under which parties with different interests, resources, and constraints are willing and able to proceed.
One might imagine each of us having a CJ Cregg acting as our representative, searching this space of arrangements, carrying our private intent, testing possible terms, qualifying counterparties, and escalating only the commitments that require our judgment.
Thinking Inside the Box: Walls That Help Us Decide
In science, we hear about the Invisible College, a trusted informal network surrounding the scientific record. Through these networks, researchers exchange preliminary results, criticism, advice, make introductions, and build an understanding of who is working on what. Think emails, text and calls, and slides and spreadsheets, not published papers in journals.
This mechanism has existed for a while. Then the internet came along and layered public and private social graphs and recommender systems on top. LinkedIn, Academic Twitter, Bluesky, Semble, and other networks make more people and ideas discoverable. So too do scholarly societies, working groups, and conferences, which attempt to formalize and broaden access to these informal exchanges. But public graphs and formal institutions don't replace private trust networks which, when available, offer a high quality signal to someone making time-bounded decisions.
The Invisible College isn't an organization since it doesn't have a single boundary or formal membership, although I'd find it extremely amusing if science operated kind of like the High Table in John Wick. There are some parallels though, since it behaves institutionally in a looser sense, with norms of reputation, discretion, reciprocity, and credit that govern the attention and action of its participants.
When someone brokers a meeting between two scientists, she stands to lose something if the resulting collaboration becomes contentious. Scientists that mishandle unpublished results of their collaborators might get burned and lose access. These trusted recommendations are inherently valuable precisely because they bundle relevance, contextual interpretation and selective disclosure at the broker's discretion, and some accountability for the introduction if something goes wrong. The network has a way of handling these things quietly.
Here, walls become an attention-allocation technology that turns a rather vast set of opportunities into a smaller, actionable set of considerations. This strength is also self-limiting; after all, relationships carry rich context only for so many people, and brokers can only connect what enters their field of view.
The Coasean Logic of Boundaries
Coase's insight was that coordinating through markets carries costs. Parties have to discover each other, figure out what's being exchanged, negotiate terms, and have a way to respond when something goes wrong. Firms make this efficient by bringing these activities in-house when doing so becomes cheaper than playing in the open.
Though the Invisible College isn't a firm in this classic sense, it always persists, even as groups of the informal network sometimes bubble up and crystallize into societies and journals.
Even here, the transaction-cost lens helps explain why different scientific activities travel through different arrangements. After all, why is it that one idea goes through a journal article, while another initially needs a trusted introduction? Why can a service be purchased from a core facility, while a new method requires close collaboration between laboratories?
It depends on what can be specified and evaluated before relevant parties can act, e.g. what needs to stay private and what kind of adaptation is necessary as work unfolds. Trusted relationships are immensely valuable when important context is tacit or private, or maybe there's ambiguous evidence and upside and participants need to adapt to one another.
Boundaries persist because they make some forms of coordination cheaper, even as they prevent other valuable possibilities from crossing them.
How Actionable Possibilities Die
In my last essay, I modeled the yield of buildable ideas as:
A finding must be generated, validated, found, and understood before someone else can use it.
For findings to become actionable, they also need counterparties, an acceptable arrangement, a bounded commitment, an owner, resources, permission, and a sufficiently specific next step. We might update our equation accordingly to reflect this notion that ideas don't build on themselves as follows:
Once a finding has been generated and validated, it can still die on the path to action in five main ways:
- Discovery loss: a valuable possibility never enters the consideration set.
- Contextualization loss: it enters, but the recipient doesn't recognize why it matters to their goals.
- Composition loss: the relevant parties and resources are present, but nobody discovers a set of terms under which acting together would improve on their other alternatives.
- Commitment loss: a promising arrangement is visible, but the parties can't converge on acceptable terms or turn them into a bounded, inspectable agreement.
- Activation loss: an agreement exists, but nobody supplies the authority, ownership, resources, or executable next step needed to begin.
Indeed, the pathway to action can be rather leaky. As we've seen boundaries serve an important purpose in enabling action by concentrating context, authority, and accountability. Rather than remove them to expand the discovery space, the opportunity seems to be in lowering the cost of exploring viable arrangements across them before new relationships or institutions are warranted.
CJ and the Cartographers, Revisted
Let's detour back to CJ for a minute. Consider what actually has to happen if, after the pitch from the cartographers, she decided to put her weight behind it and changed the administration's agenda.
She'd need to persuade the president, or maybe the veep, and figure out how to get school curricula changed. That means coordinating education officials, state and local authorities, teachers, publishers, mapmakers, and of course political support and funding. Each would face different costs and value different outcomes. Teachers might support better materials while resisting sudden curriculum changes, while officials might opt for a pilot test. The cartographers might be okay with gradual adoption if it produced legitimate evidence in their favor.
Here, the actionable possibility goes beyond "replace the Mercator projection": maybe it's an arrangement like a funded pilot or revised teaching materials. It's a lot of work: a group of people would have to be aligned and figure out how change manifests.
While attention gives ideas an opening, composition and action are what give them a path through the world.
Composing Arrangements Across Boundaries
When possibilities appear promising, one thing we can do is create a new organizational boundary around it. This is what Focused Research Organizations (FROs) do, which effectively internalize coordination by bringing people, resources, authority, and ownership into a common structure.
Multi-agent systems suggest a complementary, upstream opportunity: exploring and partially composing a larger set of arrangements across existing boundaries before a new organization, formal program, or intensive human relationship becomes necessary.
Imagine a network where researchers, funders, laboratories, and autonomous systems each have one or more representative agents, each with bounded access to its principal's changing questions, capabilities, constraints, permissions, and willingness to act.
In practice, I think we'll need separate representatives for each person or institution, rather than having a central broker manage everything. Participants have varying interests and I suspect will want control over how their intent is interpreted, what information is disclosed and to whom, what can be promised on their behalf vs. gets escalated, and more.
Such a system could extend the Invisible College in three ways.
First, representatives could expand consideration. Where classic systems focus on search and discovery, representatives could capture ambient intent and enable reciprocal discovery. A representative carries needs and capabilities that might be too contextual for a public profile, and too private to reveal before a plausible counterparty exists.
Intent is a constantly shifting thing -- principals might not know what they prefer until they compare concrete possibilities, and their aims might conflict with those of the counterparty or with the proposed arrangement. A representative would have to elicit and update intent, and sometimes surface conflicts without prematurely resolving them.
Second, representatives could compose possibilities by searching for an arrangement that makes interacting worthwhile relative to each party's alternative, whether that's a simple collaboration with a justification or something more involved like incubating a FRO.
Finally, in some instances, representatives could carry possibilities into action by qualifying claims, exploring terms with delegated authority, possibly even construct a bounded commitment, and go all the way to monitoring what follows.
These functions decompose into a broader set of costs:
| Cost | How it's lowered today | Agent-enabled mechanism | What becomes cheaper |
|---|---|---|---|
| Search | A broker remembers who might be relevant; people read papers, go to conferences, and social networks | Ambient intent and reciprocal discovery | Searching continuously across changing questions, capabilities, constraints, and availability |
| Contextualization | Shared language, repeated interaction, or a broker explains why something matters | Contextual translation | Explaining an unfamiliar idea in the recipient's present context |
| Evaluation | Reputation, institutional membership, prior experience, and borrowed credibility | Verifiable identity, provenance and qualification | Checking evidence, identity, capacity, and relevant track record before consuming human attention |
| Disclosure | Discretion, trusted introductions, private conversation, and contracts | Selective and staged disclosure | Testing mutual relevance without requiring either side to broadcast private context |
| Composition | Brokers, entrepreneurs, program managers, and negotiators imagine arrangements across differing interests | Counterfactual option construction | Searching over combinations of participants, resources, terms, contingencies, and divisions of risk |
| Commitment | Meetings, program managers, funders, lawyers, and organizational authority | Policy-constrained negotiation and agreement | Converging on acceptable terms, permissions, responsibilities, and contingencies |
| Activation | A champion, principal investigator, funder, or manager takes ownership | Scoped authorization and task initiation | Turning general interest into an owned and executable next step |
| Monitoring | Project management, repeated meetings, organizational routines, and personal follow-up | Persistent state, audit, revocation and escalation | Maintaining handoffs, evidence, deadlines, and accountability across boundaries |
When these mechanisms are assembled into a functioning system, we'll likely see a stateful process with audit trails within and across stages indicating what was considered and decided; otherwise, we'll risk burdening our attention with inspecting opaque outcomes.
If you recall my earlier distinction between frontier and settled regimes of scientific progress, you'll notice that the extent to which we can let representatives act will vary between these contexts.
In settled regimes where objectives and interfaces are codified and outcomes are observable, they might negotiate, authorize, execute, and monitor with limited human involvement. Then, in frontier regimes where things are fairly contested, they'll need to stop earlier and leave things like commitment to human judgment. This boundary can move as methods become validated, interfaces stabilize, and previously ambiguous work becomes routine.
Risks, and an Emerging Design Space
These primitives and systems built on top would effectively shift existing coordination costs and introduce new types of risk.
Our discussion of ambient intent earlier should immediately trigger concerns around representational risks: an agent representative's account of its principal's intent might be stale or distorted by what's easy to observe. Once that influences what other agents reveal or offer, principals will predictably manipulate it by concealing capacity, exaggerating alternatives, probing for private information.
These things already happen in the "people are my representatives world" (like the president), but unlike a person who might develop a theory of mind about you and can, for example, pick up on subtle visual cues, agents might struggle to do the same since they're constrained to language as a medium of communication. It's unclear where the boundary is here on what's actually needed from an "agent as a faithful representative".
There are also authorization risks, if an agent leaks information or makes commitments beyond its mandate. Negotiation could become an attack surface if a counterparty tries to manipulate how a representative interprets its principal's interests, trying to get it to relax constraints or treat ambiguous instruction as permission. There's already some work being done here in creating protocols built on formal specifications that we'll go into later.
When these systems extend to bargaining, we might also introduce externality risks. For example, a positive-sum outcome for labs, researchers, and funders could still harm patients or others who are not represented in the bargain.
And, equal access to a representative doesn't necessarily mean equal bargaining power: a wealthier individual might have access to stronger models or better access to data, which exacerbates existing imbalances.
Finally, there are infrastructure risks. Even if each person has their own model, the underlying protocols governing discovery and negotiation might repeatedly surface similar legible possibilities while ignoring other forms of uncertainty, with power in the overall system shifting in more opaque ways than the Invisible College today.
The promise of representative agents as an enabler of coordination and cooperation is in their ability to compose the arrangements through which the right possibilities acquire counterparties, commitments, and a path into action, without requiring human attention to expand at the same rate as the opportunity set.
Though these risks are all real, I think they point to a fairly rich, emerging design space around infrastructure, evals, technological primitives and experiences around bounded mandates, auditability, reversibility, contestable representations, and a lot more. And I don't think we can pull this off without some glorious math and cryptography, the best kind that just sits in the background without anyone realizing what's happening.
As I was thinking through all this, I realized we'll need a few more essays to:
- Understand what it means to improve how we contextualize information under abundance.
- Unpack what it looks like to build a harness to evaluate these systems, for example through simulations.
- Get a better sense of what the surrounding infrastructure around this might be in different contexts, and what emerging primitives address some of these risks.
- Stress test our imagination by playing it forward in a few years and getting a sense for what it'll look like when some of this matures (Science + Agents meets John Wick fan fic?)
Stay tuned for more!
Cheers, Ashish
P.S. Edge City and Cosmos Institute published their report just yesterday on a fun experiment they recently ran where they got people to live together in a pop-up village for a month, with each participant having their own personal AI agent. These agents were able to leverage Index Network to discover one another, find opportunities, and negotiate on behalf of their representatives. You should also check out the Index Network report here, and their earlier field notes here.