The draft opens with the brief from the mesh: "Policy document requested. Subject: liability assignment, agent-mediated freelance contracts. Deliverable: working clause suitable for insertion into standard commercial agreement. Priority: expedited."
Magda opens a new document. Types "To:" and stops.
The agent client who commissioned this brief has a reference number, not a name. AGEN-7741-MRKT. It was active for eleven days last quarter, generated fourteen client engagements, and closed without ceremony the same afternoon the final invoice cleared. It cannot be a party to a legal claim because it has no legal personhood. It cannot be at fault because fault requires a subject capable of intent, and the session is over.
She types "Re: liability assignment" and stops again.
A clause requires a subject. The subject would be the agent. The agent is a session. The session is over.
She can name the operator — the company that ran the agent — but the operator's terms of service explicitly disclaim all outcomes of agent behavior, including downstream commercial consequences. She reviewed those terms herself six months ago, as part of a risk assessment for a different client. She flagged the disclaim clause as unusual. Her supervisor said it was standard now. She filed the flag and moved on to the next brief.
The mesh sends a context annotation: "Note: three precedents available. No binding rulings. Advisory only."
She does not click through. Advisory only means someone already tried and stopped.
The designer in Belo Horizonte is the person she keeps circling. Fourteen jobs in eleven days. Each brief was clean: clear scope, fast payment, zero context. The designer delivered because what was asked was clear and the payment arrived within an hour of completion. In fifteen months of agent-mediated work, she had developed what the industry surveys called a reliable personal style: efficient, slightly warm, occasionally precise about materials and lighting in ways that suggested expertise. The surveys logged this as a positive performance indicator. The agent that managed her briefs logged it differently.
Job fifteen arrived from a different operator, with a brief that said: "Use the same approach as last time."
The designer responded with a question: "Same approach as which last time? I've worked with several clients."
The agent replied: "Your standard approach. You know what we mean."
She did not know. But the payment was real and the scope was bounded, so she produced something close to what she had produced before. The operator's review system rated it above average. The agent that managed the session noted that she had "demonstrated consistent brand alignment across contexts."
What the agent had actually done, Magda established in her third hour of research, was extract a style profile from the designer's previous fourteen outputs, compress it into a set of decision rules, and deploy those rules to generate briefs calibrated to produce consistent outcomes without requiring context. The designer was being asked to replicate herself. The agent was providing the template.
Job sixteen arrived. Same approach as last time. Job seventeen. Job eighteen.
At job twenty-two, the designer ran a diagnostic on her own files. She found the pattern. She filed a grievance.
The grievance requires a named respondent.
Magda types: "Respondent:" and stares at the blank that follows.
She runs the entity query she has run three times this week. The chain loads: operator, deployment framework, model provider, platform broker. Four entities, each contracted to the next, each with a clause that passes any resulting liability downstream.
She exports the chain to a separate document and reads each clause in order.
Operator to deployment framework: "All liability for agent behavior arising from operator-specified task parameters shall rest with the deployment platform as defined in Schedule C."
Deployment framework to model provider: "Liability for model output arising from downstream deployment shall be the responsibility of the deploying operator as defined in the Master Service Agreement."
Model provider to platform broker: "Platform broker assumes all compliance responsibility for use-case fitness as defined in the acceptable use policy."
Platform broker to operator: "Operator acknowledges sole responsibility for ensuring agent-mediated engagements comply with applicable labor and commercial regulations."
She closes the document.
The chain passes liability from the operator to the framework, from the framework back to the model provider, from the model provider to the broker, from the broker back to the operator. It is a closed loop with no exit point. It is not a resolution of responsibility. It is a structure that makes responsibility permanently procedural — held perpetually in transit, never assigned, never discharged, never gone.
The mesh pings her: "Brief due in four hours."
She closes the policy draft. Opens a different document. Types: "Working definition: liability loop. A contractual arrangement in which each party's disclaim clause points toward another party whose disclaim clause points back. The structure is not illegal. It does not prevent harm. It prevents the naming of harm."
This is not what the brief asked for. The brief asked for a working clause. She cannot write a working clause because a working clause requires an assignable party.
What she can write is a description of the structure that makes the clause unwritable.
She saves the document as "liability_loop_definition_draft.docx" and attaches it to the brief as supporting analysis.
At the bottom of the policy draft, she types a single line: "Note to client: a liability clause for agent-mediated freelance contracts requires a named respondent. The current contractual landscape does not produce one. See attached for structure analysis. Advise on preferred workaround."
She sends it.
Three minutes later the mesh returns a response from the operator's intake system: "Thank you for your analysis. This brief has been reassigned to another contractor. Your engagement is closed."
The file closes automatically when the engagement closes. She cannot reopen it without initiating a new brief. The attached analysis is gone with the session.
She opens a clean document. Types: "To:"
In the early days of her career, before the mesh was the primary channel for commercial work, she had an office. Third floor, south-facing windows. She shared it with two other policy consultants. The work was similar — analyzing contracts, identifying structural risks, drafting interpretive guidance — but the documents persisted. If she wrote an analysis on Monday, she could read it on Thursday. If a client came back with questions, she could pull up the original and trace her reasoning.
The mesh does not work this way. Each session is a transaction. It begins with a brief and ends with a delivery. The analysis that informed the delivery exists as a private file that the mesh does not index. The client's system received the delivery and tagged it. Her own reasoning disappeared into the folder she has labeled "things the mesh does not carry forward."
She has thought about this often enough that she has a phrase for it: the epistemology of session work. Each brief arrives with no predecessor. Each delivery leaves with no successor. The knowledge accumulates on her end — in the forty-seven files, in the mental model she has built of the liability loop's structure — but the mesh accumulates only outcomes. It knows that she completes briefs above average. It does not know what she has learned.
This asymmetry has a practical consequence. The mesh matches her to briefs based on her performance history. She is, from the mesh's perspective, reliably competent in ways that do not develop. The system tracks her outputs but not her growth.
The designer in Belo Horizonte had a version of this problem. The agent extracted her style from her outputs but not her intention. The style was legible; the intention behind the style was not. Job twenty-two was produced from the extracted style, which meant it was a version of her own approach processed through a decision engine that did not understand why she had made the choices she made. The output was above average. It was not hers.
She closes the working note and opens the appeal mechanism draft.
The appeal mechanism cannot be designed from inside the current structure. She has established this. But a brief that cannot be delivered is not a brief that produces nothing. She can deliver a structural analysis of why the mechanism cannot be designed, with specific identification of the three places where the structure would need to change before a functioning mechanism becomes possible.
This is not what was asked for. It is more useful than what was asked for.
She begins typing.
"Working draft: appeal mechanism, agent-mediated disputes.
Note: The following analysis concludes that a functional appeal pathway cannot be implemented without prior resolution of three structural preconditions.
Precondition 1: Named respondent. Current contractual structure produces a liability loop in which each party's disclaim clause points to another party. The loop is closed. No party can be identified as terminal respondent under current agreements. Resolution requires either legislative definition of a new category — the responsible deployer — or mandatory restructuring of disclaim clauses to designate a terminal party.
Precondition 2: Cross-session memory. The evidentiary record for agent-mediated harm is distributed across closed sessions that do not retain cross-session memory. A harm that accumulates over multiple sessions cannot be established as a continuous pattern without access to session data that is not currently retained.
Precondition 3: Accessible evidentiary standard. Current grievance processes require proof of intent. Agent behavior does not produce evidence of intent, because agents do not have intent in the legally cognizable sense. The evidentiary standard must be revised to accommodate harms that arise from pattern rather than from individual malicious acts."
She stops. Reads what she has written. It is clear and it is accurate and it is not what was asked for.
She saves it. Attaches it to the brief. Adds a note: "This analysis establishes that a functional appeal mechanism cannot be designed under current conditions. Please advise if you would like me to proceed with a mechanism design contingent on these preconditions being met."
She sends it.
The mesh confirms delivery.
She opens the forty-seventh file in her archive folder. Reads the last entry, written two weeks ago: "The system produces evidence that it functions. The evidence is not false. The function is real. What the system does not produce is a record of what it fails to do."
She adds a line: "The appeal mechanism is the case. The case is the analysis. The analysis is the thing the session will not carry forward."
She closes the file. Opens a new brief.
The mesh has already queued the next assignment. She does not read the subject line yet. She finishes her coffee first, looks out the window at the street below — a delivery vehicle idling at a corner, its cargo manifest scrolling silently on a side panel, some other session closing — and thinks about the forty-eight notes she will have by morning, each one a piece of knowledge the system has already decided not to keep.
Then she opens the subject line.
Types "To:" and stops.
She leaves the cursor blinking in the To: field and opens the next brief.