The Cost of Getting Things Done: Coordination Technologies in an Age of Agents
Companies and contracts are mechanisms to facilitate the coordination of human activities amid uncertainty. As agentic AI systems significantly reduce the costs associated with coordination, other unresolved challenges become increasingly apparent.
Every institution is, among other things, a solution to a problem of trust under uncertainty.
A company exists because it is cheaper, in some range of circumstances, to bring activities inside a single hierarchy than to coordinate them through a series of market transactions. A contract exists because two parties want to commit to a future course of action, and need some legal / financial / reputational mechanism to make that commitment credible. Ronald Coase's influential insight, sharpened by Oliver Williamson and others over the following decades, was that companies and contracts are not natural categories but technologies: tools for solving the recurring problem of getting parties with different information, different incentives, and different time horizons to act together in pursuit of something neither could accomplish alone.
Drug development is an unusually comprehensive test for this kind of coordination. A single therapeutic can pass through a dozen organizational forms before it reaches a patient: academic lab, commercial sponsor, contract research organization, manufacturing partner, regulatory body, payer, provider. Each transition is governed by some combination of contract, license, regulation, and informal norm. The cost of making these transitions work is not incidental to the cost of a drug. In many cases it is the cost of the medicine. A great deal of what looks like “the high cost of drug development” is, on closer inspection, the accumulated cost of coordination: the searching, negotiating, drafting, monitoring, and enforcing that has to happen at every step in the development process.
This is a lens through which we can consider a question that is at the moment more often discussed in terms of productivity than in terms of institutional structure:
What happens to coordination technologies (i.e., companies, contracts, and the arrangements built on top of them) when agentic AI systems become active participants in the coordination process itself?
Anatomy of a Transaction Cost
Williamson's framework breaks the cost of coordinating across an organizational boundary into a few recurring components.
There is the cost of search—finding the right counterparty, the right dataset, the right manufacturing slot, the right collaborator with the right expertise.
There is the cost of negotiation and contracting—translating a desired arrangement into language precise enough to be enforceable, and reconciling the inevitable differences between what each party wants and what each party is willing to put in writing.
And there is the cost of monitoring and enforcement—verifying after the fact that the parties are doing what they said they would do, and having some recourse if they are not.
In biomedicine, each of these costs is elevated by the field's particular characteristics. Search is hard because the relevant expertise is narrow and distributed—a biospecimen repository, a regulatory pathway precedent, a manufacturing capability for a specific modality, each may exist in only a handful of places worldwide, and finding them is itself a specialized skill. Negotiation is hard because the underlying science is uncertain, which makes it difficult to specify contingencies in advance—a licensing agreement for an early-stage asset has to anticipate outcomes that the scientists themselves cannot yet predict. And monitoring is hard because the things being monitored (e.g., data quality, protocol adherence, manufacturing consistency) are often only legible to people with deep domain expertise, which makes oversight difficult and expensive.
The historic response to all of this has been to build institutions that absorb these costs once and amortize them over many transactions. Law firms specializing in life sciences licensing exist because they have internalized through repetition a body of contractual language and precedent that would be prohibitively expensive for any single party to reconstruct. CROs exist because they have built monitoring infrastructure that individual sponsors would find wasteful to duplicate. Standard-form agreements (such as material transfer agreements, clinical trial agreements, and data use agreements) exist because they encode in advance the results of thousands of prior negotiations such that future parties do not have to renegotiate from scratch.
What all of these elements have in common is that they are solutions to the problem of expensive intelligence-based coordination applied to recurring situations.
And expensive intelligence coordination applied to recurring situations is increasingly one of the key things agentic AI systems are good at.
What Agents Are Actually Good At
It is worth being precise about what changes with agentic systems and what does not, because the temptation in discussions of AI and institutions is to leap directly to grand restructuring—the end of the firm, the rise of the DAO, contracts that execute themselves on a blockchain—without first establishing the more modest claim on which any of that would have to rest.
The modest claim is this: agentic systems lower the cost of search, negotiation, and monitoring as activities, independent of who is performing them or what organizational form surrounds them.
An agent that can read a thousand licensing agreements, a regulatory database, and a set of internal capability records should be able to identify a plausible counterparty for a given collaboration far faster than a business development team working through its network. An agent that can hold the relevant precedent, the other party's prior agreements, and the current draft in working memory simultaneously should be able to iterate on contract language at a pace no human team can match. An agent that can continuously read data feeds, compare them against protocol specifications, and flag deviations should be able to monitor a clinical trial or a manufacturing process in something closer to real time, rather than at the cadence of quarterly audits.
None of this requires the agent to have authority—to be a signatory, a fiduciary, a party to anything. It requires only that the agent be a capable participant in the expensive cognitive work that currently sits inside law firms, business development functions, and compliance departments. This is a much narrower claim than “AI will run companies,” and it is also likely a much more certain one.
The question worth asking is not whether this will happen, as it seems inevitable, but rather what it does to the shape of the institutions that were originally built around the very costs such systems will reduce.
The Boundary Of The Firm, Revisited
To begin to think about institutional shapes in an age of agents, start with the most basic question: if agents compress the cost of transacting, does the firm itself shrink?
Coase's original question, why does the firm exist at all rather than everything being conducted through market contracts?, has a famous answer: because the cost of using the market (search, negotiation, contracting, monitoring for every single transaction) can exceed the cost of bringing these activities inside a single hierarchy where they can be directed by command rather than negotiated by contract. The boundary of the firm in this view sits wherever those two cost curves cross.
If agentic systems lower the cost of market transactions without correspondingly lowering the cost of hierarchy, the boundary of the firm should shift toward the market. Activities that were brought in-house because having to contract externally for them every time was too expensive, become contractable again. We might expect to see more finely grained specialization: smaller entities, organized around narrower capabilities, transacting with each other more often, because the overhead of each individual transaction has fallen.
But the symmetry is not obvious, and this is where the analysis has to resist a tidy conclusion. Agentic systems might lower the cost of internal coordination just as much as external coordination. Monitoring within a single firm can after all face the same problems as monitoring across a contractual boundary, and an agent that can read internal data feeds in real time could reduce the cost of command-based hierarchies too.
If both internal and external costs fall in roughly equal measure, the boundary of the firm does not move much; it simply becomes cheaper to operate on either side of it. And there is a further possibility worth taking seriously: that the costs which most determine the firm's boundary in biomedicine specifically are not the transactional costs that agents are best at reducing, but the costs of trust—of being willing to share a biospecimen, a dataset, a piece of unpublished science with an external party at all. Those costs are not primarily informational. They are costs of exposure, of reputation, of what happens if the relationship goes wrong. An agent that can draft a perfect data use agreement in minutes does not, by itself, make an institution more willing to share its data; it just means that if the institution becomes willing to share (trust), the agreement is no longer the bottleneck.
This suggests a useful distinction. Some of what currently sits inside firms, or inside elaborate contractual structures, sits there because of transactional friction—and that friction is now compressible. Some of it sits there because of something else: incentive misalignment, risk allocation, the desire to maintain optionality, or simply the fact that trust between organizations takes time to build regardless of how cheap the paperwork becomes. Agentic systems will be very good at telling these two categories apart in the sense that they will rapidly eliminate the first kind of friction while leaving the second kind exactly where it was.
The interesting question is what the resulting landscape looks like once that sorting has happened.
The Contract As A Living Object
There is a second-order effect to agentic systems that is somewhat separate from the firm / market boundary question but that is worth dwelling on: what happens to contracts themselves when the cost of monitoring approaches the cost of reading.
A contract is, among other things, a bet on the future made under conditions where continuous re-negotiation is too expensive. Parties specify terms in advance (e.g., milestones, royalty tiers, data-sharing obligations, termination triggers) precisely because sitting down to renegotiate every time circumstances change would consume the value the agreement was meant to create in the first place. The static, point-in-time nature of contracts is not an essential feature of agreements; it is an artifact of the cost of negotiation.
If that cost falls substantially, the contract starts to look less like a fixed document and more like a standing relationship with adjustable terms—something closer to a protocol than a one-time agreement. A clinical trial collaboration might specify not a fixed data-sharing schedule but a set of conditions under which data-sharing terms are automatically reviewed; a licensing agreement might specify not a single royalty structure but a mechanism by which royalty terms adjust as new information about the asset's value becomes available, with agents on both sides continuously evaluating whether the current terms still reflect the deal both parties thought they were making.
This adjustable construct is not a new idea; option structures, milestone-based payments, and re-opener clauses are all attempts to build adjustability into agreements without paying the full cost of renegotiation every time. What changes is the granularity at which adjustability becomes economical. Re-opener clauses exist for the handful of contingencies important enough to negotiate explicitly in advance. An agentic system makes it economical to monitor for, and propose adjustments around, contingencies that no one thought to negotiate in advance at all, because the cost of noticing that circumstances have changed, and drafting a proposed adjustment, has fallen below the threshold at which it is worth a human's time even to consider whether the contract still fits.
The risk, of course, is that an agreement which can be continuously reopened is not really an agreement in the sense that gave contracts their value in the first place—the value of commitment, of being able to plan around a known set of terms. There is a version of this future in which agentic contracting makes every relationship more responsive and more efficient, and a version in which it makes every relationship perpetually unsettled, with each party's agents continuously probing for terms slightly more favorable than the ones currently in force. Which version predominates will depend less on the technology than on the norms, and eventually the rules, that govern how and how often agentic renegotiation is permitted to occur. That is a governance question not a technical one, and it is the kind of question that tends to get settled by precedent and convention long before it gets settled by explicit policy.
What This Does Not Change
It is worth closing where Mancur Olson and Elinor Ostrom would want us to close: with the reminder that lowering the cost of coordination is not the same as solving the problems that coordination was meant to address.
A great deal of what makes biomedical collaboration difficult is not that the relevant agreements are expensive to draft or monitor, but that the parties' interests are genuinely misaligned, and no amount of contractual sophistication changes that. A pharmaceutical company that does not want to share trial data with competitors is not failing to share it because the data use agreement is too costly to negotiate; it is not sharing it because sharing is, by its own lights, against its interest. An agent that can produce a perfectly calibrated, continuously adjustable, fully monitored data-sharing agreement in seconds does not change that calculus. If anything, by making the transactional cost of sharing approach zero, it isolates the incentive cost more clearly than before. It now becomes harder to point to friction as the reason something didn't happen, and easier to see the underlying choice for what it is.
This is a way to hold the two halves of this essay together. Agentic systems are likely to be genuinely transformative with respect to the transactional layer of coordination in biomedicine (and elsewhere)—search, drafting, monitoring, the continuous low-grade cognitive labor that currently absorbs so much of the cost of doing science and medicine across organizational boundaries. These systems are much less likely to be transformative with respect to the governance layer—who gets to set terms, whose interests get represented, what counts as a fair allocation of risk and reward. If anything, as the transactional layer becomes cheap, the governance layer becomes more visible, and more consequential, by comparison.
The coordination technologies of the next decade may look less like better contracts and more like better answers to the much older question of why the parties to those contracts want what they want in the first place.♦
