June 13, 2026 · 5 min read

My AI Sandbox.

Everything I recommend to a client, I have broken myself first — in my own sandbox for principles, in the team’s real environment for architecture.

The sandbox has one job: find where an idea breaks before a client pays for it.

Two testing grounds

Testing happens in two places. They answer different questions.

My own sandbox is the home model. Here I test principles: how a pattern behaves, what it does to a service and the people inside it, where it saves time and where it produces confident nonsense. It is fast and disposable, so I can try ten versions of an idea in an afternoon.

The team’s environment is where principles meet the organisation. We work with operational data, safely, inside the systems the team already runs. The architecture conversations happen here, against the systems the organisation actually runs on. Together we map what the team can build in low- or no-code, what needs custom development, and how to connect agents to real systems so they can act on a user’s behalf.

The home ground tests the principle. The real environment tests how it best connects to the operational reality.

Both grounds run the same two frames: a person thinking with a model, and a person directing several. I call them Sparring Spaces and Agent Fabric.

i. Sparring Spaces — the assistive phase

The model works as thinking partner, challenger, and reviewer. The human stays the author. What turns one good session into a better next one is the memory layer underneath.

Functional view · 01Assistive phase
Human holds intent, judgement & authorship — stays in the loop
sparring
Model · postures
Partnergenerate
Challengercritique
Reviewerevaluate
the working loop
Reason·Act·Remember
FrameDraftChallengeReviewDecide
Workingshort-termthis thread · scratchpad · live decisions
Recalllong-termEpisodic past sessions & what held up · Semantic domain context · Profile how I think
Functional lineage — human-in-the-loop & reflection; memory after CoALA (Sumers & Yao). After Mollick (Co-Intelligence), Ng (reflection), Weng (agent memory).
The assistive phase: the human keeps authorship and judgement; the model supplies generation, critique and review as named postures and runs a reason–act–remember loop; a two-tier memory — working memory plus episodic, semantic and profile recall — carries continuity across sessions.

Fig. 1 — A sparring partner earns the name only if it remembers what was already tried.

A sparring space is working when the model disagrees with you on the record and the disagreement is still there next week. Memory is what makes the challenge real: the model can say you tried this in March and it did not hold.

I prove the loop on my own ground first. Then the team runs it on their real work: their documents, their cases, their operational data. That second run is the one that counts, because it tests the pattern against the organisation’s real material.

ii. Agent Fabric — the agentic phase

Here a person directs several agents that reason, call tools, and remember, inside explicit boundaries. A commission carries an objective and success criteria. Tools sit behind a common protocol. Checkpoints decide what a human sees before anything becomes real. Trust starts at the boundary.

Functional view · 02Agentic phase
Director human — sets the commission, owns the outcome
objective + success criteria
Guardrails
Orchestrator reason → act → remember · decompose · delegate · synthesise
Agent
Agent
Agent
⊞ MCP tools / APIs▤ Memory · short + long
short agent state · messages  |  long semantic · episodic · profile · prompts
Evaluator · checkpoint pass → synthesise  ·  fail → loop ↻  ·  breach → escalate ⤴
BoundaryIdentity·Fine-grained access·Full audit trail
Functional lineage — agent = LLM + planning + memory + tools; tools via MCP (Anthropic, Model Context Protocol). After Weng; Anthropic (orchestrator–workers, evaluator–optimizer); memory after CoALA.
The agentic phase: a bounded commission flows down through an orchestrator that reasons, acts and remembers; worker agents share MCP-exposed tools and a short- and long-term memory; an evaluator gates every handover, and a hard boundary of identity, fine-grained access and audit holds underneath — anything past the guardrail escalates to the accountable human.

Fig. 2 — Every arrow that crosses the guardrail is a decision someone has to be able to answer for.

Trust is earned at the evaluator checkpoint and the boundary under it: who gets escalated to, on what breach, with what audit trail. Build the guardrail first.

The agentic phase is where the work moves into the team’s environment early. An agent acting on a user’s behalf has to touch real systems: the CRM, the case system, the calendar. So we map which systems expose usable interfaces, what the access model lets an agent do, and where the low-code platform hands over to custom code. Those conversations happen when the people who run the systems are in the room.

A leader needs to know what breaks first, what their teams can build themselves, what they will have to commission, and what they are accountable for once it runs without them in the room. The sandbox answers that before it costs anyone anything.

Andreas Conradi · June 2026