The industry spent one year on prompts, the next on context, and has spent this one discovering that what actually determines whether an agent is useful lives in neither. It calls the missing thing the harness. It calls building the thing harness engineering, and describes it as the third phase of a maturity curve, arriving after the other two have exhausted themselves. The framing is correct as far as it travels. What it does not see is that the question sitting underneath the harness was never an engineering question. It is a question about the conferral of authority on a party whose fitness you have not established, and it is old.
The word for it is adhikara. From adhi and the dhatu kri: authority over, the fitness to undertake, the eligibility to receive. In Panini’s grammar an adhikara sutra is a governing rule whose jurisdiction runs forward across the sutras that follow it - a rule whose whole content is the scope of its own authority. In the pedagogic tradition, adhikara is the answer to a question the modern academy has forgotten how to ask: not whether a student can understand the teaching, but whether the teaching, entering this particular student, will do what it is supposed to do. Krishna closes the Gita with the injunction. Do not speak this to one without tapas, without devotion, to one who does not wish to hear. Vedanta hardens it into a checklist and puts it before the inquiry rather than after: viveka, vairagya, the six disciplines, the burning desire for freedom. Acquire these, then we begin. This is routinely read as gatekeeping by people who assume all withholding is status protection. The tradition’s actual claim is narrower and more interesting. A teaching received by a vessel unfit for it does damage, and the damage is not primarily to the vessel.
The Wrong Axis
Everyone working with agents is now conducting the same experiment, and almost everyone is measuring the wrong variable. The question in the air is how capable the model is. Benchmarks, evals, the endless comparison of one release against another. Capability was never the binding constraint, and it has not been for some time. The binding constraint is adhikara, and adhikara has a property that capability does not: it is domain-specific and it does not transfer.
This is the part the general-intelligence discourse cannot process. A grammarian of the first rank has no adhikara in ritual. A man who can recite the whole Ashtadhyayi is not thereby fit to conduct a rite, and no amount of additional grammar will make him fit. The fitness is local to the domain and is established inside it, by demonstration, under observation. Capability generalizes. Authority does not.
So the agent that writes flawless React has no adhikara over your billing migration, and its performance on the React work tells you nothing about the migration except that the underlying shakti is adequate. Shakti was never in doubt. Every deployment failure I can identify reduces to the same shape: authority granted disproportionate to fitness demonstrated in that specific domain. Write access to production granted on the strength of good test-writing. Autonomy over a refactor granted on the strength of good code review. Irreversibility granted on the strength of reversible successes. The operator generalized from capability, because capability is what he was measuring, and capability was the wrong axis.
The agent arrives with enormous shakti and zero adhikara, and unlike a student it does not accumulate adhikara across the discontinuity between one session and the next unless you build the accumulation yourself.
That inversion is the whole novelty of the situation. In the classical relation the shishya arrives with little capacity and builds both capacity and standing over years, and the two grow together closely enough that the guru can use one as a proxy for the other. With an agent they have come apart completely. Full power on day one, no standing ever, and a memory that ends when the window does. We have never had to confer authority on anything shaped like this.
The Instrument Is the Restraint
The harness is a yantra, and the word is not a decoration on the argument. Yantra comes from the dhatu yantr, which Panini’s list glosses as sankocana - restraining, curbing, contracting, checking. The most literal rendering of yantra is an instrument of restraint. A bow is a yantra because it restrains the string. The chariot’s tackle is a yantra because it restrains the horse. The geometric yantra of the tantric traditions is an instrument for restraining diffuse divine potency into a form a human being can actually work with.
Set this against the Greek. Mekhane, from mekhos, means contrivance, expedient, device, means of getting around. The machine in the European inheritance is a trick played on a limit. The god descends from the crane because the plot has run out of legitimate moves. Two etymologies, two complete philosophies of technology: one in which the tool is the cheat, one in which the tool is the discipline.
Heidegger saw far enough into the first to name it. Das Gestell, the enframing, the disposition that sets upon the world and holds it as standing-reserve, and his correct insight that the essence of technology is nothing technological. But he had only two positions available to him, because Greek tekhne gave him only two: enframing on one side, poiesis and the bringing-forth on the other, with the whole hope of the essay resting on a return to the second. He had no third term. The Indian root supplies it. The instrument as restraint is neither a setting-upon nor a bringing-forth. It is a narrowing of the possible that makes a particular actuality reachable, and it is the only thing that ever converts potency into work.
Which is why the industry’s ledger on this is upside down. Sandboxes, read-only mounts, required tests, schemas on tool inputs, permission prompts before writes, the whole apparatus of constraint gets booked as safety tax - capability that existed and was then subtracted for prudence. Unbanked water is a flood. Banked water is a river, and only a river irrigates. Nobody describes the bank as a tax on the water. A model of unbounded capability with no yantra around it produces nothing you can ship, and the reason is not that it is unsafe. The reason is that a flood does no work.
The practical consequence runs against instinct. When an agent underperforms on a task, the reflex is to widen its permissions and enrich its context so it can find its own way. The move that works more often is to narrow the aperture until the task has one shape, and then let full power operate inside that shape. The narrowing is the engineering.
Pramana in the Graph
The live finding of this year is that verification, and no longer generation, is the bottleneck. Agents write thousands of lines of working code and the question becomes whether anyone can audit them before they run. Correct diagnosis. The remedies on offer mostly consist of helping humans verify faster, which is a treadmill, and the wrong layer.
Nyaya gives the right layer. A pramana is a means of valid knowing, and the school’s discipline is that every claim must be traceable to one, and a claim that cannot be so traced is opinion wearing knowledge’s clothes. Four of them: pratyaksha, direct perception; anumana, inference from a mark; upamana, knowing by comparison; shabda, valid testimony. The last carries a condition that the others do not. Testimony is a pramana only when the speaker is apta - trustworthy in the strong sense of possessing direct knowledge of what he reports. Testimony from one who lacks independent access to the fact is not weak evidence. It is not evidence.
Now put every node of an agent graph to the test. A node with a real test suite has pratyaksha: something outside the model perceived directly whether the thing worked. A node behind a type checker or a schema validator has anumana: valid inference from structure to conclusion, cheap and sound within its scope. A node with a well-built eval has a coarse pratyaksha. And a node whose only validation is the model’s own report that the output looks correct has shabda alone, from a speaker who is not apta about his own production, because he has no access to the world independent of the production. That node knows nothing. Its output has to travel.
Every node either carries its own pramana or exports its verification burden to the root, and the root is one human skull with a fixed daily bandwidth.
This is the actual reason parallel fan-out stops paying. The field has converged on three to five parallel subagents as the practical ceiling and explains it by coordination overhead and context loss, which are real. Underneath them is arithmetic. Ten agents produce ten times the artifacts and ten times the audit, and if none of the ten can prove itself, all ten audits land on the same person, who could only ever verify three. The other seven get waved through, which is theatre, and the system has now converted compute into unreviewed change at a very high rate.
So graph engineering has exactly one question in it, asked at every node: can this node prove itself, or must its output reach me? Count the nodes that cannot prove themselves. That integer, and not the model’s score on anything, is your throughput. Every hour spent giving an internal node a pramana of its own buys more than any hour spent on the prompt above it. This is what people are reaching for when they say the tests matter more now than they used to. The tests were always the pratyaksha. What changed is that the volume of claims needing one went up by two orders of magnitude.
What Survives the Night
The agent has no continuous memory. What persists across the discontinuity is only what you inscribed: the skill file, the plugin, the instruction living in the harness, the conventions document the thing reads on waking. These are samskaras in the exact technical sense - latent impressions laid down by past action that condition future action without being recalled as memories. In the Indian analysis of mind, samskaras are what make a person a person rather than a sequence of unconnected mental events, and the jiva carries them across the discontinuity of death itself. The agent’s discontinuity is smaller and arrives far more often, several times a working day, but it is the same problem and it takes the same solution.
Writing a skill is therefore not documentation. It is samskara-nirmana performed on something that would otherwise be reborn ignorant every morning. And the corollary is where almost everyone loses the compounding: the highest-leverage hour in the whole practice is the one after the work, spent converting what you just learned about how this thing fails into an impression that will condition it tomorrow. Most operators spend zero hours there. They correct the same failure daily, at a rate of once per day forever, and describe this as working with AI.
The impression has to encode the failure, not the success. Success is generic and the model already possesses it. An instruction saying to write clean, well-tested code is inert; the weights contain that and more. The impression that carries information is the one local to your terrain: that your migrations must be reversible and the ORM’s generated down-step is wrong for enum changes, that this service’s retry logic is load-bearing and looks redundant, that the last four times an agent touched the auth middleware it broke the same edge case in the same way. That is a samskara. The other thing is a mantra chanted at a machine that has already memorized it.
Sraddha and the Rubber Stamp
What becomes of the human in this? Two postures dominate, and both are failures of the same faculty.
Sraddha gets translated as faith, which imports the wrong apparatus entirely. Its sense is closer to a placing of the heart: a provisional extension of trust that lets inquiry begin, since no one has ever learned anything from anyone without extending some. The tradition never lets it terminate in itself. Nyaya insists every claim reduce to a pramana. The Buddha tells the Kalamas not to go by report, by lineage, by scripture, by inference, by the seeming competence of the speaker, nor by the fact that the monk is their teacher, but to know for themselves. Vedanta makes hearing only the first of three movements and puts manana, the reasoning through, immediately after it. Sraddha opens the door. Verification walks through it.
The first failure is andha-sraddha, blind trust, the approval reflex. The operator who accepts every diff has not delegated, he has abdicated, and the loop now contains no adhikari at all - not the agent, which never had standing, and not the human, who has stopped exercising the judgment that was his entire contribution. The work is unowned. Run this for a year and you get a codebase no living person understands, which is a codebase nobody can change, which is a dead codebase that happens to still return 200.
The second failure is the operator who reads every token, who cannot let a subagent run without watching the stream, who reviews the plan and then reviews the execution against the plan. He extended no sraddha and so he received nothing back. He has an expensive autocomplete and eyestrain, and he is now performing both the work and the supervision of the work, which is strictly more labour than doing it alone.
The posture that works is the guru’s, and the tradition specifies it closely. Give work at the edge of present adhikara. Check the result against a pramana you control rather than one the student supplies. On the strength of that check, extend the adhikara by an increment, in that domain only. Revoke on failure. Upanayana and the staged release of texts are exactly this technology - a protocol for handing over authority in portions matched to fitness demonstrated under observation. We built it for beings who took twelve years to complete the sequence. We are now applying it to something that completes it in an afternoon and forgets by evening, which changes the clock speed and changes nothing else.
What holds for one operator and his agent holds for the collective and its institutions, because an institution is precisely a body granted adhikara to act on behalf of people who cannot verify its work directly. Every institutional catastrophe has the identical shape: authority exercised well past demonstrated fitness, with the verification loop that was supposed to bound it long since decayed into ceremony. The legislature that votes on the bill it has not read. The regulator staffed from the industry it regulates. The peer review that reviews nothing and confers everything. Andha-sraddha at civilizational scale, and the same arithmetic as the ten unverifiable subagents, running for decades.
We are about to build institutions out of agents and grant them authority over things that matter. The discourse asking whether they are conscious, whether they are general, whether the curve is exponential, is a distraction from the only question that has ever governed the handing over of power. It is the question the householder asks before giving up the keys to the granary and the one the guru asks before speaking the difficult verse. What has this thing demonstrated, in this domain, and by what means would I know if it were wrong.
Not how much it can produce before breakfast. Anything can produce. The Kaurava army was also large.