Intent-Conditioned Task World — The Temporary World an Agent Actually Thinks Inside

Longer context windows make it possible to dump an organisation’s knowledge into a model. They do not make that wise. What an agent actually thinks inside is a temporary, purpose-shaped world — and that object can be designed.

The demo always starts the same way. Someone points an agent at the company drive, enables retrieval, and asks for strategy, critique, or a deck. The model is fluent. Citations appear. And the work still reads like a clever outsider who skimmed the intranet on the way in from the airport.

That failure is rarely a shortage of tokens. It is a shortage of world.

Models do not reason from the organisation’s entire graph. They reason from whatever has been made attention-resident for this task: pages loaded, edges followed, doctrines co-present, absences named or ignored. Open-ended knowledge work improves when the durable wiki is compiled into an ephemeral, provenance-bearing intent-conditioned task world — shaped by intent, role, date, access, and attention budget — because that temporary object is the cognitive environment the agent actually inhabits.¹

Not the wiki, not the window

Three objects get confused in almost every agent conversation:

1. The durable wiki — the organisation’s semantic fabric: claims, entities, typed relationships, supersession, provenance.
2. The active context window — what currently occupies model attention (CPU cache, not warehouse).
3. The task world — the temporary, intent-shaped projection compiled for one act of cognition.

Boot profiles choose where an agent enters the fabric. An intent compiler turns a parent purpose into probes and fused coverage. Bloom raises resolution where independent routes converge. None of those mechanisms is the full temporary object they jointly produce. The task world is that complete temporary epistemic object: selected regions, joined interpretations, dissent preserved, known absences listed, sources receipted, expiry declared.²

The wiki is the durable world model. The task world is the working model assembled for this particular act of cognition.

Epistemic conditioning, not “find a fact”

Retrieval is usually sold as evidence supply: look up a passage, paste it, answer. Something more powerful happens when curated organisational material loads before the agent solves. What the model reads changes what it notices, which analogies become available, which options feel plausible, and which options it refuses to propose.

In strategy work, loading a deployment doctrine before evaluating a customer opportunity can make the agent refuse the obvious customer-facing proposal — not because a citation forbade it in the last paragraph, but because the doctrine changed the shape of the proposed strategy. That is retrieval-as-stance: the frameworks condition posture before solve.³

Call the positive mechanism epistemic conditioning. Security prompt injection is untrusted material illicitly crossing into the instruction layer. Epistemic conditioning is curated material deliberately shaping understanding. The token physics is similar; trust, purpose, and architecture differ. Three intensifying forms:

• Map injection — orient the agent to the intellectual territory that exists.
• Retrieval as stance — load doctrines that alter posture before solving.
• Attention-residence — keep those ideas active across the hundreds of micro-forks for which no separate lookup would ever be issued.⁴

That is why “the smartest answer in the room” understates the product. What you can convene is the best-prepared room the organisation has ever put around one worker for one purpose — fragments of people, projects, and lessons that have never previously shared a single working memory.

Four layers, one fabric

“One wiki or many?” is the wrong binary. The mature stack separates persistence and purpose:

Sales and delivery should not maintain rival copies of “Project Orion was cancelled in month seven.” They should share the canonical event and hold lens-qualified implications: do not sell factory throughput before staffing evidence exists; reuse assumptions need project-specific validation; double-staffing seniors produced burnout risk; productisation reduced near-term billable contribution before margin improved. One organisational memory, many disclosed lenses, many temporary task worlds.

A companion piece on the executable worldview covers how that activated world joins authority, action, and outcome write-back. This article stops at the temporary object those runtimes inhabit.¹

Make the task world first-class: the manifest

Walks already build temporary maps implicitly — in tool traces, pages opened, bloomed regions, remaining context. The next step is an explicit task-world manifest:

Parent intent
Role and declared lens
As-at date
Access scope
Boot profile
Questions / probes generated
Pages and edges included (and why)
Convergence hotspots
Important minority paths
Counter-lens
Known absences
Sources opened
Derived interpretations
Expiry / promotion policy

That is more than a search result set. It is a temporary model of “the organisation relevant to this purpose,” inspectable before and after the agent works. If you cannot answer why a region was included, what was deliberately left out, and when this world expires, you do not yet have a designed object — you have a lucky context dump.

Proof shape: two intents, one graph

Minimum proof is not a larger model. It is two materially different intents over the same durable fabric producing two manifests that share canonical truth and diverge in lens and task-local additions.

Illustrative. Intent A: prepare a go-to-market critique of a proposed expansion. Intent B: prepare a delivery-risk review of the same expansion. Shared: project history, cancellation precedents, capacity facts. Divergent: sales implication edges vs delivery and people overlays; different minority paths; different known absences (pricing experiments vs staffing evidence). The agent that inhabits World A proposes different micro-moves than the agent in World B — without either inventing a private “truth” about what happened.

An ablation completes the picture: generic similarity retrieval vs boot-profile entry alone vs full task-world conditioning (probes, bloom, absences, counter-lens, attention-residence of doctrines). Expect the third condition to change refusals and micro-decisions, not only which quotes appear in the appendix.

What faster compute changes — and does not

Cheaper cognition can move more interpretation into runtime: a previously untouched folder can be read just-in-time under acquisition-fit, integration risk, workforce, concentration, and regulatory probes. Those interpretations may have little value outside the task, so pre-persisting them would be waste.

Faster compute does not remove the durable fabric. Even a cheap just-in-time compiler still needs canonical entities, established relationships, prior contradictions, supersession, vocabulary, and lessons already integrated. Otherwise every task reinterprets the organisation from raw bronze and never compounds.

Why ordinary agent memory is the wrong centre

Most “agent memory” is episodic: what I recently did, what the user said, what worked last time, a conversation summary. It is first-person, recent, agent-scoped, and accumulates forward from zero. Useful as a staging buffer. Disastrous as institutional truth.

Directly promoting recent experiences and prior model outputs into durable worldview produces recency bias, self-confirming summaries, accidental policy formation, compounding hallucinations, and one agent’s interpretation masquerading as the organisation. Agent memory can die with the agent. Institutional memory must survive it — observed, evidenced, decided, contested, superseded, learned from outcomes.⁵

Threat and governance (sketch)

• Access inheritance: task worlds must not widen the user’s reach; compile under declared access scope.
• Untrusted content: bronze and third-party text condition cognition only through typed interpretation layers — never as raw instruction.
• Provenance: every included region carries why and from which sources; derived material ranks below bronze-backed claims.
• Expiry: default death of the temporary object; no silent immortality of a lucky walk.
• Promotion: only evidence-backed residues enter fabric or lens overlays under review — never “the agent said it once.”

Epistemic conditioning is power. Governed, it gives knowledge workers a co-present organisational field. Ungoverned, it is a sophisticated way to launder untrusted text into posture.

Operate the lifecycle

After this argument you should be able to:

1. Define a task world as distinct from wiki, window, boot profile, and walk trace.
2. Compile one from parent intent, role, date, access, and attention budget.
3. Inspect the manifest: inclusions, hotspots, minority paths, absences, sources.
4. Expire it when the act of work ends or the as-at date invalidates it.
5. Promote only residues that earn durable status under explicit policy.

Long context removed the old excuse that “we cannot load enough.” It did not remove the design problem. The design problem is task-shaped epistemic construction: building the temporary world an agent actually thinks inside — then letting it die cleanly so the durable fabric keeps compounding.