Agent-Ready Knowledge Architecture
I build the systems organisations run AI agents on.
Agents are only as good as the memory they run on and the workspace built around them. I architect that whole layer: the sources of truth, the context budgets, the loops agents run in, and the checks that keep them honest.
Free tools
Instruments from the working system. Everything runs in your browser.Workspace Maturity Check
Eighteen questions on how your agent workspace actually runs. Banded score across six dimensions, gaps ranked, fixes linked to the reference architecture.
Run the check →Agent Readiness Assessment
The organisation view. Twelve questions on whether your organisation is ready to put AI agents to real work, and which exposures to close first.
Assess readiness →Redaction Pre-Flight
Paste what you're about to publish. Catches emails, home paths, credential shapes and secret filenames before they ship. Your text never leaves the tab.
Check a draft →Context Carry-Cost Calculator
What a big early read really costs across a session: size times turns remaining, priced. Seeded with measured numbers from my own workspace.
Price a read →Learn
Two tracks through the same architecture — the modules that explain why, and a tutorial that has you build one yourself.
Start learning →01 The work
Most AI adoption stalls because agents have nothing reliable to run on.
The models are strong enough. What's missing is the memory underneath them. An organisation runs on a context layer that mostly isn't in tidy systems: stale documents indistinguishable from current ones, decisions buried in inboxes and chat threads, processes that run on tacit rules nobody wrote down, no source defined as authoritative. Point a capable agent at that and it retrieves confidently from the wrong place.
My work is that layer and the machinery around it. Three things separate a demo from a system an organisation can actually run on. The workspace architecture: the governed memory, roles, task flow, and guardrails a fleet of agents operates inside. The token economics: what a capable agent costs per unit of work, and the orchestration and model-tiering decisions that hold that cost down as usage compounds. And the loop engineering: which work runs unattended, which gets surfaced for human review, and which stays manual. I build all three, on a memory system designed to be read by agents as much as people.
02 Capabilities
Three primary practice areas
Each one runs daily in a production workspace and is documented in the open, pattern by pattern, in a public reference architecture.
AI workspace architecture
The full environment a fleet of agents operates in: governed organisational memory, specialised expert roles, task and delegation flow, hook-enforced guardrails, and a self-audit that keeps the whole thing honest. Designed so a small team, or one operator, can run many agents without losing the thread.
Proof in the open: agent-workspace-architecture — a redacted production workspace with 18 expert roles, an adoption guide, and an interactive tour.
Token & context optimisation
Context is a budget, not a constant. Per-run cost ledgers, budget caps on scheduled agents, cache-aware context design, bounded fan-outs, and tiered model execution: plan at the frontier tier, execute a tier down, review at the top. The discipline that decides whether an agent system stays affordable as usage compounds.
Proof in the open: the reference architecture's token-budget module and its pattern “Context is a budget, not a constant”.
Loop engineering
Which work belongs in an autonomous loop, which gets surfaced for human review, and which stays manual, decided by test rather than enthusiasm. Scheduled lanes with freshness sentinels and a dead-man's switch, delegation queues an operator fills deliberately, and verify-loops that only stop on a checkable condition.
Proof in the open: the reference architecture's loop-selection and delegation-queue patterns, distilled from retiring an always-on agent that a gated queue outperformed.
Backed by
Evaluation & verification
Golden-set regression suites, LLM-judge design, adversarial verification passes, and synthetic canaries in a weekly self-audit. A capability change ships with a measurable signal behind it, or it doesn't ship.
Agent governance & safety
Permission floors, hook-enforced guardrails, untrusted-content handling, credential discipline, action logging, and human-review gates. Autonomy sized to evidence, expanded only by explicit grant.
Knowledge & memory architecture
The practice core: source-of-truth mapping, context packs per workflow, and memory governance with provenance, retention, and typed memory that points at sources instead of copying them.
03 Selected work
Built, not demoed
A redacted, world-readable snapshot of a production agent workspace: eighteen load-bearing patterns, an adoption guide, samples, and an interactive tour. The architecture, inspectable in the open.
Run daily in production on a designed memory system: persistent cross-session memory with consolidation discipline, a source-of-truth pointer architecture, decision and lesson logs, and an automated self-audit. 18 specialised roles operate over it. The entire architecture is mapped on one page.
How organisations re-architect their knowledge and operations for a workforce of agents: the move from scattered documents to a governed memory agents can be trusted to run on.
04 Background
An engineer's discipline, applied to organisational memory
I came to knowledge architecture through data engineering, building production data platforms: cloud data warehouses, ETL pipelines, and the analytics executives actually decide on. Then I went deep on agent systems, building a full multi-agent workspace from the ground up, shipping an AI SaaS end-to-end, and formalising the work through doctoral research into how autonomous agents share and govern memory. The two threads meet at the same place: the system of record an organisation runs on, now designed for agents to read as well as people.
It means I bring an engineer's production discipline (version control, testing, provenance, rollback, observability) to a field that is still mostly demos.
- Doctoral research — trustworthy shared memory for collaborative AI agents (James Cook University, from 2026)
- Master of Data Science — James Cook University
- Graduate Diploma in Research Methods — James Cook University
05 How we work together
Engagement models
Discrete, scoped engagements with a deliverable that ships, not open-ended retainers.
Agent-Ready Knowledge Audit · 4–6 weeks, fixed scope
Where organisations start. A source-of-truth map, a context risk register, an agent-readiness scorecard, the priority use cases, and a 30/60/90 roadmap.
Agent workspace & cost review · 2–3 weeks, fixed scope
For teams already running agents. Your workspace architecture, token and context economics, loop design, and evaluation coverage, reviewed against a production reference architecture, delivered as a prioritised fix list with the cost case for each change.
Context architecture build · 8–12 weeks, one workflow deep
Stand up the context layer for one high-value workflow: source register, business glossary, decision memory, context manifests, retrieval rules, output templates, and the evaluation tests that prove agents use them.
Agent context assurance · regulated & government teams
For teams that must show their agents use trusted sources: evaluation cases, provenance and citation behaviour, sensitivity handling, and records, audit, and compliance alignment. Scoped and fixed-price.
Strategic advisory & mentoring sessions
Short-form architecture and strategy sessions on workspace architecture, token economics, loop engineering, memory governance, and AI-first organisational design — for technical leaders, founders, and architects building this layer inside their organisations. Direct enquiry, or via expert networks.
How every engagement runs
- Bounded scope — one high-pain area at a time (a key-person departure, an onboarding bottleneck, a "nobody can find anything" domain), never "map our whole organisation".
- Non-extraction by design — analysis computes in your environment and emits derived structure, not content; raw material never leaves your tenant, and the data flow is auditable in your own logs.
- Auditable over autonomous — anything that runs is transparent, scoped, throttled, and logged, deliberately less autonomous than the technology allows; access expands only by explicit grant.
- An explicit IP split — findings about your organisation are yours; the generalised methods I arrive with and refine are mine. Written into every engagement, never discovered later.
I build the capability, then hand over the keys. The goal is to make the engagement unnecessary, not to create a dependency.
Not a fit: generic ML or analytics builds, staff augmentation, or open-ended on-call retainers. I focus on the knowledge and context layer agents run on, in engagements that ship and end.
06 Work together
Available for select fixed-scope engagements.
Your agents are only as good as the mess they read. If that's the problem worth fixing, get in touch.
Or follow the work as it ships — patterns, teardowns, tools, and essays on agent workspaces:
Writing Agent-First, a book on re-architecting organisational knowledge for a workforce of agents. Get notified when it ships.