How the exchange stays alive.

The evidence on knowledge platforms in this sector is blunt, and uncomfortable: they fail on community design, facilitation and curation — not technology. PLACE Australia — the $38.62M national intermediary for place-based change — has had a Knowledge Hub publicly "in development" for 20 months with no visible platform; the sector's previous attempt sits dormant. So an approach paper that leads with AI agents owes the room an immediate answer to the obvious objection: isn't this technology pointed at a community problem?

This paper is the answer. The agentic layerAI agents working under human direction is not the product, and it is not a substitute for the community model. It exists to resolve the contradiction the evidence leaves open: these platforms only live when someone tends them continuously, and the same evidence says the sector systematically under-resources that tending — 16% of online communities now run with zero full-time staff, the highest on record (CMX industry report, 2025), and buyers in this market evaluate maintenance burden as sceptically as build cost. The approach keeps the human judgement that makes an exchange trustworthy, and uses agents to make that judgement affordable at the scale a national exchange needs.

What follows was developed and deliberately pressure-tested in a structured working session on 12 June 2026. The pre-mortemimagining the project has already failed, then working backwards to the causes at section 8 came out of trying to kill it. The companion opportunity case carries the full market and demand evidence; this paper carries everything needed to judge the approach itself.

4 truths about knowledge in communities

4 pillars — a theory of why these platforms die

A living exchange continuously works 4 pillars. Platforms die when any one of them goes unstaffed — and the "digital bookshelf" every practitioner recognises is simply what remains when all 4 stop.

The second hypothesis answers PLACE's declared agenda rather than ours. PLACE's energy is publicly flowing to impact measurement, not the knowledge hub — and the approach turns that from a threat into the strongest card in the deck:

Section 5 develops this in full. The point to hold here: H1 makes the exchange affordable to run; H2 makes it worth funding. Together they let us pitch the same artefact as community infrastructure to the field and as sensing infrastructure to the funder — without either framing being spin.

The gardener-amplifier model

At the centre of the exchange are the gardeners — anywhere from 1 to 5 trusted humans — whose judgement the agents amplify. The gardener is a role, not a person: the exchange launches with at least 1 named gardener and is designed to grow toward a bench — a PLACE practice lead, practitioners from the field on fractional hours, a curator embedded in a community — different people tending different corners of the garden. Agents work the routine continuously: scanning for decay and disconnection, suggesting links, drafting welcomes and targeted contribution requests, flagging gaps and demand. The gardeners review, steer and overrule through evaluation dashboards built for the purpose — not moderators buried in queues, but curators with instruments. A named lead gardener is accountable for everything the agents do; and where an agent's work reaches outward — a welcome, a named contribution request — the agent drafts and a gardener sends: nothing leaves the system under automated initiative.

The direction of growth matters as much as the floor. The agents exist to complement human judgement, never to thin it out: when the budget grows, it buys more human gardening before it buys more automation. The sustain envelope (section 4) prices gardening hours, not headcount — the same money can fund 1 full-time gardener or 5 people tending a day a week each. More hands is the better design, because trust travels with the people who hold it: gardeners embedded in different corners of the field hear more, and are trusted by more of the community, than any single curator could be.

This makes at least 1 named gardener a precondition of the engagement, not a hope. If PLACE cannot name its first gardener, the first piece of work is designing and standing up the role — and we say so rather than building around the vacancy. (The pre-mortem names what happens otherwise: the phantom gardener, section 8.)

2 trust surfaces, 2 mechanics

The community is the facilitator

The deepest version of Relation is not agent-to-person — it is agents brokering person-to-person: finding the practitioner who has done the thing, and making the introduction. The agent is never just serving documents; it is translating in both directions — research into a practitioner's context, practitioners' problems back to the field — and handing off to a human the moment a human is the better answer.

The practitioner-facing test of all of this is deliberately unglamorous: alive is the only metric that shows. Practitioners never see the architecture. They see whether the exchange feels tended, current and responsive — or like every platform before it.

The innate line: deterministic clockwork

Most of what keeps an exchange feeling alive is not intelligence — it is reliability. Link integrity, decay scans, demand signals, rubric checks, scheduled digests, escalation triggers: all of it runs as deterministic, scheduled clockworkcode that produces the same output for the same input, every time — no AI model involved — auditable, testable, near-zero marginal cost, incapable of hallucinating. The first design question for any capability is never "which model?" — it is "does this need a model at all?"

The adaptive line: intelligence on escalation

Inferencerunning an AI model to produce an output — the metered cost of using AI is reserved for what genuinely needs judgement under ambiguity: drafting a connection between 2 pieces of knowledge, summarising a story for a different context, holding a conversational search, proposing who should meet whom. Above the adaptive line sit the gardeners — the escalation point for anything novel, sensitive or uncertain. The system's centre of gravity stays cheap; intelligence is spent only where it earns its cost.

The rest of the machine, named

The agents earn the attention, but they are perhaps a third of the build. The rest is named here because anything trust-bearing that lives only in prose ends up unpriced — and the systems below are where most of the responsible-AI posture is actually enforced. Each gets a name and a job at this altitude; the technology choices behind them are sprint work, not approach work.

Stage 1 inherits rather than introduces: PLACE's identity, hosting, security boundary and support arrangements stay exactly where they are, with the exchange running inside them. That is the architecture's answer to the IT-partner veto (section 8) — there is no new estate to refuse.

Cheap models, made deterministic

We start with the cheapest models that pass the evaluationsscripted test suites that check an AI system's outputs against agreed pass criteria — before launch and continuously after — including open-source — and engineer reliability upward: constrained outputs, narrow tasks, evaluation gates, deterministic scaffolding around every model call. Model capability keeps rising and prices keep falling; an architecture built on cheap-models-made-reliable rides that curve, while an architecture built on premium models inherits someone else's pricing decisions. Platform and provider selection is treated as a long-term token-economics decisionwhat a vendor charges per unit of AI usage — the recurring price of every model call, compounding over years, and the blueprint stays platform-portablethe design can be rebuilt on a different vendor's platform without starting over. This is what makes the $100–200k/yr sustain envelope honest: the envelope prices the tending as a whole — the gardeners' hours, the agents' run cost, and the hosting and support underneath them — and the mix flexes with usage and budget, so the exchange survives funding cycles instead of dying with its first grant. The envelope is a design target, not a quote; validating it against real usage is part of what the sprint is for.

The safety model is reversibility, not perfection

We do not promise agents that never err — we promise an exchange where no agent action is final: attributed, logged, one-click reversible, and explained in the product (the commitments below). The architecture also names its own failure mode — autoimmunity, the system attacking healthy knowledge — and designs tolerance against it, with the gardeners as the check. When something does fail, the failure is encoded as durable memory attached to the knowledge itself, so the next community inherits the lesson rather than repeating the mistake.

The automation boundary map

Which tasks are clockwork, where agents may infer, and where humans decide is not an emergent property of the codebase — it is a designed, reviewable artefact, co-created with PLACE, held by a small governance group — PLACE, practitioners and community representation — and revisited as trust grows. This is the deepest sense in which the agentic layer is design work and the platform is commodity: the boundary map, the rubric, the interaction constitution and the trust surfaces are judgement calls about a community, made with the community. The build is downstream of those calls — which is why the engagement leads with design, and why the platform underneath it stays interchangeable.

6 architectural commitments

PLACE's declared agenda is measurement — and its own thesis is that change appears first in relationships, trust and practice, long before population data moves. Nothing in its measurement marketplace is built to observe that layer. A working exchange can — as a by-product of doing its job.

Every time the exchange brokers an application — a practitioner finds a practice, adapts it with help, and reports how it went — that is a practice-diffusion data point: what travelled, where, into what context, with what result. Today this layer of change is invisible; it lives in phone calls and hallway conversations. An exchange that helps practitioners apply knowledge is, structurally, the first instrument that can see it — a sample of the layer, not a census of it. The data point exists only when a practitioner closes the loop and reports back, and designing for that return is part of the sprint's work.

3 properties make this more than a dashboard feature. First, the loop is the sensor: applicability can only be measured through continuous engagement — a survey can't do it, a repository can't do it, only an exchange practitioners keep returning to can. Second, the instrument appreciates: every month of operation makes the diffusion record more valuable, which means delay has a real cost — an argument that reverses the usual procurement instinct to wait. Third, the measurement schema itself is treated as living content: the sector's frameworks evolve, so what the instrument measures must be re-measurable rather than baked in.

Governance is single-flow by design: engagement data flows only to PLACE, for its own impact reporting — we design the sensor; the data and the story it tells are PLACE's. Structurally, the diffusion record is its own store — held apart from the knowledge graph and from practitioners' session context — and it crosses to PLACE through a consent gate: collected with practitioners' knowledge, aggregated before it travels. Single-flow is a stance, not a solved problem: it is exactly where the consent question below begins. Section 7 sequences when this instrument switches on.

Quality is not recency. A 2019 story that matches a practitioner's context beats this quarter's report that doesn't. The exchange is built on a different quality metric — applicability — and it changes what the agents actually do.

Applicability is the quality metric

Because applicability is relational — quality lives in the match between a story and a context — the agents' centre of gravity is matchmaking, not janitorial work. And because staleness is really disconnection — decay lives in a piece of knowledge's connections, not its publication date — the curation verb changes: less pruning, more weaving. An agent's first response to ageing content is to test and renew its relations, not to flag it for deletion.

Rubric over gatekeeper — worn on the outside

Quality assurance is a rubric, not a gatekeeper: completeness across visible dimensions (context, evidence, outcome, who-to-talk-to), applied consistently by agents. The rubric is worn on the outside — friendly, visible flags that read as contribution prompts ("this story would travel further with an outcome note") rather than admission gates. Intake stays generous; curation happens on the way out, exactly as practitioners in the field insisted. And the rubric protects voice rather than standardising it: texture, dialect and first-person grit are a protected class, because a beige exchange is a dead one.

Red links, stories, and the contributor's return

The knowledge graph advertises its own gaps: where practitioners keep searching and finding nothing, agents surface the missing node and turn it into a targeted, named contribution request — Wikipedia's red links, pointed at proven demand. Contributions travel as stories, because the practitioner journey is find → consume → relate → apply, and narrative is the vehicle of relatability. And contributing pays a visible return: discoverability, beacon status on a topic, and — the part no platform has ever delivered — lessons flowing back from other communities' applications of your story.

2 hard problems in this section stay named rather than hidden. Failure vanishes by default — nobody publishes what didn't work under their own name — so capturing it is an explicit design problem (anonymised aggregation, the intermediary as publisher of record), and captured failures become the durable memory cells of section 4. And small communities are identifiable by detail: contributed stories need a light sensitivity tier, which is open design work, not a solved checkbox (section 9).

The sequence is not a delay on the exchange — the first risk to this opportunity is that the window closes quietly. The sprint and the internal deployment are how the window is held: they put us inside the relationship while the evidence is built, and the exchange's design work runs inside them, not after them.

"Why Not PLACE?" — the framing that carries it

The pitch frame is sector leadership, not efficiency: organisations spending public and philanthropic money have a defensible obligation to take AI's leverage — done safely, governed visibly, and stated openly. For PLACE — whose whole mandate is showing the sector how to work better — being the first to do agentic knowledge infrastructure responsibly is on-mission, not a risk to it. That makes the responsible-AI posture itself a deliverable — observabilitylogging and monitoring that makes every agent action inspectable after the fact, evaluations, governance, staff fluency — and opens a natural ongoing role: fractional AI-enablement alongside the build. That role is recurring revenue shaped like capability building — our existing engine.

The commercial honesty

The moat is not the idea — everything in this paper could be photocopied. The moat is the earned position: field knowledge, enabled champions, and trusted-partner standing that a competitor would need a year inside the sector to replicate. Against the key-person risk this creates, the work is spec-first and person-independent: definition before delivery, a clean approved specification as the asset, AI-assisted build with contractor capacity in place. And the blueprint stays platform-portable — PLACE is never locked to a vendor, including us. Exit is designed, not asserted: open-format export of the graph and the record, PLACE owning the spec and the data, and a named archival home for the stories if the funding ever stops — because a paper whose whole thesis is that platforms die owes the sector an answer to where the knowledge goes when this one does.

We ran the approach through a deliberate sabotage exercise: 8 credible ways this fails, each now countered by a design decision you have already read. The counters are in the approach because the failures were named first — not the other way round.

3 questions are genuinely open. They are carried as discovery questions in the sprint — named in the room, never papered over — because in this sector, candour is the relationship.

Next moves

Nothing happens before the Place-based Impact Summit (15–16 June 2026) — deliberate patience is the move; PLACE's agreed actions and post-summit momentum are the entry conditions the 90-day plan is built around. This paper goes to the partner group alongside the opportunity case for the go/no-go. After the summit, the approach travels into the first conversation framed inside the measurement agenda: the exchange as the instrument for the layer of change PLACE's frameworks can't yet see — with the agentic layer as how it stays alive, and stays affordable.