How agents decide, when they stop, and a local story kept as evidence

Abstract: A day of articles on deciding: what a real decision needs, why an agent that can go anywhere has no reason to stop, what happens when every mistake adds a rule, and one customer service agent built three ways to count which rules are only hoped for. Alongside, a live local story, a hospital outage nobody could check from home, kept as evidence while it was happening.

What does an agent need to decide well, and what makes it stop? Most of 8 October's articles answer one of those two questions, from the person's side and from the agent's. Alongside them, a live local story, kept as evidence while it was still happening.

A decision is more than a yes

Agency is not a yes sets out what anyone asked to decide actually needs, a person or an agent: options beyond yes, context in their own terms, time, somewhere to escalate, authority over the system that produced the request, and incentives that treat a wrong yes and a wrong no alike. The scale it builds has seven levels, the weakest dimension caps the whole decision, and its vault holds the scale. One line from it is worth keeping:

Whatever you punish, you teach the other answer.

Knowing when to stop

Knowing when to stop starts with people, security champions who had automated away whole classes of bug and then debated whether GUIDs were random enough, and comes to agents, who have the same problem by design:

An entity that can go anywhere has no natural reason to stop anywhere.

Every mistake added a rule is what happens when the answer is more instructions: long sessions, prompts that grow to 100 KB, and each new rule causing the next mistake. The way back it maps is to turn each piece of the process into a small, shipped component:

If a session needs 100 KB of rules to be productive, the rules are doing the job that structure should do.

And who are you protecting against? gives security a place to stop: name the attacker first, on a six-tier ladder, and draw the line where the attacker is.

The question is who, not how much.

Hope or enforcement

The day's lead on the site is hope or enforcement: one customer service agent, one mandate, built three ways, from a single capable model connected to everything to a team of narrow agents behind a deterministic gateway. Each design gets an Agent Behaviour Policy, and the published vault counts how much of each policy is a boundary and how much is hope:

Every design has a policy for every behaviour. The difference is who enforces it.

Encrypted memory for agents that run somewhere else is the infrastructure underneath: five ways to give isolated, short-lived agents a memory that outlives them, with only ciphertext on the server.

A local story, kept as evidence

The waiting room knew first was written on the afternoon it happened. Staff at two London hospitals told patients the IT systems were down; nothing a local person could check before leaving home showed it. The article is about that gap, where local information comes from now, and the evidence behind it is kept in a vault, with care over what it does not claim:

Blocked is a limit of our access, never evidence about the hospital.

One correction from the day before: in the bridge, followed to the end, the reporter's base salary had been counted as both her income and her cost. Her row now shows pay, salary plus commission, and every other figure is unchanged.

Everything published on 8 October

Seven articles were published from 2026-10-08 to 2026-10-08, grouped below by what they are about.

Deciding and stopping

What a decision needs, and what lets a person or an agent stop.

• Agency is not a yes: a scale for human and agent decisions, from rubber stamp to the reviewer who fixes the source. A yes or no is not agency: seven dimensions, a seven-level scale for people and agents, and the review as a QA step that fixes the source.
• Knowing when to stop: what experience gives people, and what we have to design into agents. Knowing when to stop is the hard part for people and agents: what experience gives people, and the constraints that give agents the same perspective.
• Every mistake added a rule: complexity, agents, and the way back to shipping. When every agent mistake adds a rule, complexity wins: map the process, move each piece right as a small shipped component, and keep the rigour for the work.
• Who are you protecting against? Draw the security line where the attacker is, not above it. Before deciding how secure to be, name the attacker: a six-tier ladder, an air-gapped Mac mini read against it, and eight startups drawing their line.

Policy that holds

Which rules an agent keeps because it is told, and which because it cannot do otherwise, and the memory that outlives a run.

• Hope or enforcement: one customer service agent, three designs, and who keeps each promise. One customer service agent built three ways, each with an Agent Behaviour Policy: how much of each policy is hope, and what one run can reach.
• Encrypted memory for agents that run somewhere else: sgit deployment patterns, from a Mac mini to Kubernetes. Agents in isolated, ephemeral places need memory that outlives them. Five sgit deployment patterns, from a Mac mini to Kubernetes, with ciphertext-only servers.

Local, live

A story written while it was happening, with its evidence kept.

• The waiting room knew first: a live local story, and the gap where local information used to be. Staff at two London hospitals said the IT was down; nothing a local could check showed it. A live local story about where local information comes from now.

This is issue 2 of the SGit Newsroom newsletter, also published on LinkedIn in Deterministic GenAI. Every article it links to is on sgit.ai, with its sources and its data. To get the next issue by email, subscribe at sgit.ai/subscribe.