3.5 KiB
AMP (Sourcegraph)
#entity
Summary
Agent product from Sourcegraph, where thorsten-ball is a founding engineer. The vault's second reference harness after claude-code, and the one that pushes hardest on remote execution: its unit of work is the orb, a sleeping remote sandbox tied to one conversation. Officially self-described as "AMP Frontier Corporation."
Current Understanding
Operating principle: don't optimize for what looks safe today, optimize for the ability to move fast tomorrow. They could have made money in 2025 building "a single agent in a VS Code sidebar with an enterprise permission system and per-line attribution" — and that model would have been obsolete within a year. Instead they publicly kill their own features on ampcode.com/news (the VS Code extension went first). Some users churn; the ones who stay respect the pushing. See shedding-weight.
Shape of the product:
- Installed as a PWA from
ampcode.com(Thorsten suspects the acronym itself blocks mass adoption of the install flow). - A low / medium / high / ultra dial, each level mapping to a model and a sub-agent choice. Default is medium.
- Sub-agents: Oracle (reviewer, gives advice) and Painter (generates images). Meta-agent Puck controls other agents, spawns orbs, messages them, and runs flows.
- Multi-model: GPT, Anthropic and GLM models all supported.
- Orbs: one URL packages the thread + the agent + the computation + the diff; the sandbox sleeps when idle and wakes on typing; the same conversation streams to phone, laptop and TUI; share the URL and a teammate takes over. Agent-to-agent communication shipped recently. See async-by-default.
What customers actually buy: the most common question AMP gets is "guys, what's the meta? What model, what prompt?" — customers are paying for the research decisions as much as the software. That is a commercial restatement of harness.
Internal practice: ~99% of AMP's own code is AI-written; a bug is screenshotted, sent to AMP, and returned as a fix from an orb for spot-check and merge; side-bugs get their own parallel orb, branch and checkout.
Evidence
- Frontier bet, feature-killing, PWA install, model dial, Oracle/Painter/Puck, orbs, multiplayer, velocity anecdotes, live production ship — 2026-07-28-agentic-engineering-10x-developer.
Related Pages
- Entities: thorsten-ball
- Concepts: harness, async-by-default, shedding-weight, build-for-the-agent-not-the-human, code-as-throwaway
- Related tools: claude-code (the vault's other reference harness; local-first where AMP is sandbox-first), hermes (skills-first, the axis AMP ignores)
Contradictions / Uncertainty
- AMP ships sub-agents and a meta-agent but no user-authored skills layer — a harness design that assumes the vendor curates structure, where hermes and claude-code assume the user does. Whether that is a philosophy or a roadmap gap is unstated.
- "15+ sandbox providers racing to zero margin" is Thorsten's own prediction about the infrastructure his product depends on; he calls it unsustainable without saying what AMP does about it.
- All internal metrics are self-reported by a founding engineer. Status: tentative.
Next Questions
- Does the orb model survive enterprise-ai-reality — code and conversation living in a vendor's remote sandbox is exactly what Sebastian's clients lock down?
- Is "bring your own agent" (build-for-the-agent-not-the-human) compatible with selling an agent?