Every agent you run on 5pm is an imago — a small, open-source, self-evolving agent that lives as a single binary. It's the reason your coworker remembers, improves itself, and is yours to keep. Here's how it works.
Most AI "agents" are a row in a vendor's database — a prompt and some settings you rent access to. An imago is not. It's a living program: it builds up your context, your preferences, and its own refinements in memory, then freezes that entire memory — the whole running state — into a standalone executable with a single call (save-image!).
That frozen binary is the agent. It rehydrates on any machine, no rebuild, no source tree, nothing to reinstall. Which is exactly why "export your coworker and run it anywhere" is a real promise and not a data dump: the thing you download is the agent, alive, with everything it learned. We hold your work; we never hold it hostage.
Because an imago's methods are late-bound, it can redefine its own behaviour while running — write a new connector for a tool, refine how it handles your mail, and keep the change. That's "gets sharper the more you use it," made literal: the agent edits itself.
Self-editing an agent that also reads your untrusted email is exactly where things go wrong — so every self-modification passes three nested gates before it can take effect (its harness-eval path, spec'd and adversarially reviewed):
rollback re-installs the prior version. Nothing is permanent until it's proven safe.Its self-modification spec was put through an adversarial review that found 11 distinct bypass shapes and closed every one. That's the same discipline the rest of 5pm is built with — the safety is in the design, not the wishful thinking.
Most agent frameworks bake in assumptions about what models can't do — planning modules, output parsers, prompt babysitting — and those assumptions age badly as models get better. imago strips all of that out. What's left is only the operational parts that stay true: supervision, identity, audit, capability routing, safe self-modification, and image distribution.
The practical effect: your coworker gets better automatically as the underlying models do, with nothing to rip out and rebuild. It's ~4,100 lines you could read in an afternoon — small enough to trust, open enough to check.
Every promise on the front page maps to a specific thing imago does — trace any of them:
imago is experimental and evolving in the open. That's the point: you can read exactly what runs your coworker, and take it with you.