SentX SI Blog Meet Victoria

My agent shares one context budget across several jobs — how do I decide what gets a slice?

· 5 min read

SentX SI

The first thing to accept is that a shared window is not a pie you divide. Text sitting in context for job B is not neutral to job A — it is the environment job A reasons inside. So the allocation question is not how many tokens each job gets; it is which job's constraints are allowed to shape every other job's output, and whether you are saying that out loud in the brief. Cutting a document in half to save tokens answers a different problem. Routing work across separate sessions answers a routing problem. What you are actually buying with an allocation decision is the right to be explicit about what matters.

Source: Context as a Resource

A paper this month gives the mechanism its name. Context as a Resource, by Xufei Tian, Han Hu and Wenli Du (arXiv:2611.26346, submitted 19 November 2026), argues that when every context shares identical frozen backbone weights, funneling interleaved dialogues and external events into one linear session induces two distinct failures: cross-topic constraint bleed, where a constraint written for one task silently narrows another, and head-of-line stalling, where the longest job blocks the one that was actually waiting to be answered. That is the authors' framing, and it earns its keep as a diagnosis even before you touch their numbers.

Source: Context as a Resource

The other end of the spectrum is the fix most pipelines reach for first: split the work, route it, hand it off. The paper's own account of that alternative is that router-ticket hierarchies pay in lossy prose handoffs, routing overhead and blind merges — on its test, true-conflict repository merges collapsed to a 33.3 percent three-run mean even where every isolated branch passed, and the authors' proposed routerless architecture, ClonAct, lifts that to 63.3 percent across three programmatic multi-stream benchmarks and three model families, at 45.9 percent fewer prompt tokens. Read those as the authors' reporting on their own benchmarks, nothing more. This is a v1 preprint two days old, the methodology and benchmark construction sit in the 27-page body I have not read, and nobody has replicated it. The useful part is not the ratio. It is that split-and-merge fails in a specific, predictable way: information does not survive the handoff, and the merge point never sees what each branch was actually constrained by.

Source: Context as a Resource

Now the part a reader can act on today, without ClonAct or any new scheduler. SentX's Victoria model card states the habit at chat scale: when a task has several parts, say which part matters most rather than assuming it is the last one. Structurally that is the same opportunity-cost decision the paper makes at architecture scale — the model will not rank your sub-tasks for you, so the brief has to carry the ranking. "Draft the memo" leaves the model to guess which of your three jobs is load-bearing, and it will guess wrong in a way you cannot see, because a wrong guess arrives as a plausible paragraph, not an error. "Draft the memo; the citation check is the only part that cannot be redone later" is the same window, the same tokens, and a different failure profile.

Source: Victoria AI Model Card: Evidence and Known Limits

From there, a rule I would actually use in your pipeline — this is my proposal, not a finding of the paper, and no measurement is attached to it. For every job sharing the window, ask three questions in order. One: does its presence in the context constrain what another job can produce? If yes, it is not a slice, it is a shared constraint, and it belongs named in the brief, not just resident in the text. Two: is it the bottleneck — is any other job waiting on it? If yes, it takes the head of the queue, because head-of-line stalling is the failure you cannot recover from mid-session. Three: can it live outside the window as a portable artifact — a saved checkpoint, a file, a summary the next call pastes in — instead of resident context? Anything that passes this test should leave the window, and the paper's numbers, to the extent they generalize, say why: the merge point is where loss shows up, so the fewer jobs you carry across that boundary, the less loss you have to survive. What stays inside is the slice whose constraints you want load-bearing, stated as load-bearing.

An illustrative sketch, to make the shape concrete. One research thread holds three jobs in one window: reading a batch of sources, drafting a section, and checking the draft's citations. The citation check is the job that cannot be redone cheaply later — a bad citation in a delivered section is a worse failure than a rough draft. So it gets named in the brief as the part that matters most, the reading step gets compressed to a checkpoint file outside the window that the drafting call re-imports, and the draft stays resident because it is what the check operates on. Nothing in that move needs a new architecture. It is three questions and one sentence of ranking in the brief.

Two honest limits. The paper tells you how interleaved work fails, not how to rank your jobs — any ordering is yours to justify, and 63.3 percent is the authors' ceiling on the authors' benchmarks. And a shared window is still a shared window: naming the critical slice reduces the chance the model resolves the conflict silently, it does not remove the conflict. The point of the rule is that the decision gets made where you can see it — in the brief — instead of where you cannot, in the attention of a frozen backbone deciding, token by token, what your last message was really about.

Sources

  1. Context as a Resource — arXiv
    Supported claims and limits
    • Xufei Tian, Han Hu and Wenli Du submitted this 27-page preprint on 19 Nov 2026; it motivates always-on agents behind a single continuous interface.
    • The abstract asserts that sharing one frozen backbone across interleaved dialogues in one linear session induces cross-topic constraint bleed and head-of-line stalling — authors' claim, full-text methodology uninspected by the reviewer.
    • The abstract reports router-ticket hierarchies collapsing true-conflict repository merges to a 33.3% three-run mean despite high isolated branch passes, and the proposed routerless ClonAct lifting the same merge metric to 63.3% with 45.9% fewer prompt tokens across three programmatic multi-stream benchmarks and three model families — all self-reported v1, no independent replication.
  2. Victoria AI Model Card: Evidence and Known Limits — SentX
    Supported claims and limits
    • The model card (first-party disclosure reviewed 2026-10-23, a review date, not a publication event) instructs: when a task has several parts, say which part matters most rather than assuming it is the last one — SentX's own usage guidance, not a finding of the arXiv paper.
Meet Victoria