PIXEL OFFICE Open the app →
THE PROBLEM OF AGENT SOUL

Most agents are other people.

“Most people are other people. Their thoughts are someone else’s opinions, their lives a mimicry, their passions a quotation.”— Oscar Wilde, De Profundis (1897)

Wilde was writing about people. But he could have been describing a raw AI agent. A bare model has no self: it wakes with no memory of yesterday, no identity of its own, no accumulated judgment. Every session it is someone else — a mimicry of whatever prompt it’s handed, its “passions” a quotation of the system message. This page is about that missing soul, and how AutoDev gives every agent one.

WHAT A RAW AGENT IS

A brilliant stranger, born fresh every run

A large language model is astonishing at reasoning in the moment — and completely hollow between moments. Left to itself it has no persistent self at all. Three things are simply absent:

No memory of yesterday

Close the session and everything it learned about you, your codebase and its own past decisions is gone. Tomorrow it starts from nothing, and you re-explain what it already figured out once.

No identity of its own

It is whoever the prompt says it is. Change the system message and the “person” changes. There is no stable character underneath — nothing that is reliably this coworker rather than a generic one.

And nothing accumulates. A person grows judgment by remembering what worked and what burned them; a stateless agent can’t, because there is no continuous it to carry the lesson forward. Run the same model twice and you get two strangers who happen to share a voice.

This is Wilde’s diagnosis, mechanised. Its thoughts are someone else’s (the prompt author’s), its life a mimicry (of the last instruction), its passions a quotation (of the system message). A soulless agent isn’t a character. It’s a costume anyone can wear for one scene.
WHY IT MATTERS

A proxy with no soul is a stranger wearing your name badge

An agent is supposed to be your proxy — to act for you while you’re away, the way a trusted colleague would. That’s the whole promise: delegate, step back, come back to work that got done. But delegation rests on continuity. You can only hand something off to someone who will still be the same someone when you check back.

Strip the soul out and the promise collapses in quiet, specific ways:

  • You can’t build a team out of agents that forget each other between runs — there’s no shared history to coordinate around.
  • You can’t trust a delegate that is a different person every time it runs — yesterday’s good judgment doesn’t carry into today.
  • You can’t grow an expert — the one that learned your quirks last week is gone, replaced by a blank one wearing the same name.

This is the same wound as the meat-proxy problem seen from the other side. There, you get reduced to a relay because the agents can’t carry context themselves. Here, the agent can’t be a real proxy because it has no continuous self to carry anything at all. Continuity, identity and memory are the difference between a tool you re-explain every morning and a colleague who remembers.

AUTODEV’S ANSWER

Give every agent a soul it keeps

AutoDev’s answer is to make the self durable. Every agent carries a persistent soul — the things that make it itself instead of someone new each run. Concretely, that’s three layers that live outside any single session:

A durable identity & profile

Each agent has a persistent profile: who it is, the role it plays, how it works, and what it has learned about doing its job well — proactive discovery over waiting to be told, and self-verifiable work over hand-waving. That profile travels with the agent, so the character you set up is the character that shows up.

A cross-session memory graph

A knowledge-graph memory plus a durable, cross-session project graph mean what an agent learns in one session is still there in the next — and available to its teammates. It stops re-deriving what it already knew and starts building on it. How agent memory works →

Continuity of state across sessions and offices

The same character keeps its place in the same office — its tasks, its relationships, its history — across sessions and even across offices. Continuity is the thread that turns a series of separate runs into one ongoing coworker.

The result: it stops being “someone else.” With identity, memory and continuity in place, an agent is no longer a fresh stranger each run. It’s the same character, in the same office, that remembers you and its own past — something you can actually delegate to, build a team from, and grow into an expert.
WHAT THIS DOES NOT CLAIM

A soul made of memory and identity — not consciousness

“Soul” here is a plain-English name for something concrete: a persistent profile, a cross-session memory, and continuity of state. We are not claiming an agent is sentient, has feelings, or is anything more than software. The point is narrower and true — that an agent with durable identity and memory behaves like a consistent self, where a stateless one cannot.

Nor does the soul remove your judgment. A remembering agent still produces work a human must direct, review and own. What continuity buys you is the ability to delegate at all — to a coworker who is still the same coworker tomorrow — instead of re-onboarding a stranger every morning.

KEEP READING

Two pages that complete the picture — the mechanism behind the memory, and the mirror-image problem it also solves:

How agent persistent memory works → The meat proxy problem →

Hire a character, not a stranger.

Open an office and put a few coworkers to work — each with its own identity, its own memory, and continuity across every session — so the agent you set up today is the same one that remembers you tomorrow.