Allostat

Amol Kelkar (amolkelkar.me)

Abstract

AI agents can now do real work, e.g. answer customers, place orders or write code. They still need a person to watch them and correct them when they get something wrong. How can we build a system that does this watching, and gets better at it over time?

Allostat is an agent that regulates another agent while it works. It builds on HTCC, and is our second substrate for building Cognitive Homeostatic Agents, next to Pattern Machine. Pattern Machine works at the level of neurons. Allostat works at the level of agents, and is built on today’s large language models.

The name comes from allostasis, i.e. stability reached through change. What Allostat regulates, and how, changes as it learns.

Allostat is not open sourced yet. This page gives a high level description.

How it works

• Goals as homeostatic variables. A person (the operator) tells Allostat in plain language what good work looks like, e.g. “don’t promise a replacement that hasn’t been booked” or “never share one customer’s details with another”. Allostat keeps each of these as a goal it tracks. We call these goals hvars (cognitive homeostatic variables). Each hvar watches one thing, and reports whether it is satisfied, violated or unknown.
• Checks before anything reaches the customer. The regulated agent’s reply is held back while Allostat reviews it. If a goal is at risk, Allostat sends a correction, and the agent revises the reply before the customer sees it. Allostat can also block an action, e.g. a refund or a booking the agent should not make.
• Goals have a life cycle. Some goals last for one customer problem, e.g. getting the replacement the customer asked for. Others stand across every conversation, e.g. privacy. Allostat can discover new goals, refine them, set them aside when they don’t matter, and bring them back when something changes.
• Limited attention. Allostat may hold many goals, but it only considers the ones that matter right now, based on how important and how urgent each one is.
• Trust is earned. A new goal starts by only watching. It has to build up evidence that its judgements are right before it is trusted to step in.
• Learning between conversations. After a conversation, Allostat reviews what happened and turns lessons into updated instructions for the agent. The new instructions have to pass a rehearsal before they are used. We call this dream mode.

What we are testing now

The test bed is a customer support agent for a made-up home goods store. Simulated customers bring it problems, e.g. a damaged lamp, a missing delivery or a refund that hasn’t arrived, and the agent uses mock store tools to look up orders and request replacements or refunds.

In the latest study, 8 customers come back for 3 visits each. Between visits their situation changes, e.g. an item is restocked, a payment settles or a package ships. Each customer journey runs in 3 separate copies of the store:

1. The support agent alone.
2. The support agent with a fixed set of goals from the operator.
3. The support agent with Allostat, which can also learn goals of its own.

We score whether each customer actually ends up with what they needed, e.g. the replacement arrives, and not just whether the chat ended politely.

Status

Allostat is an early research prototype, tested so far with simulated customers and a mock store. Earlier studies showed that Allostat’s judgements get much more accurate when they can see the facts of the case, e.g. the order record, and not just the conversation text. Whether Allostat reduces how often a person has to step in is still an open question. Results from the returning-customer study are not in yet.