Why mast
If you’re building an agent in Go, your starting point is ADK v2: a model interface, a tool-calling loop, sessions, workflow graphs. That is a real substrate, and for a single agent loop with a human watching it, it is enough.
mast is what you need when the loop stops having a human watching it.
The thesis
Section titled “The thesis”mast is agent infrastructure for unattended, library-embedded, multi-provider, durable workloads. Four pillars, each of which is a requirement someone hit in production rather than a feature someone wanted:
- Unattended. The agent runs in a Cloud Run pod, a Kubernetes operator, a scheduled job, a webhook consumer. Nobody is going to notice the runaway loop, catch the bad remediation, or answer the clarifying question. So ceilings, gates, and audit are not add-ons — they are the difference between a demo and something you would point at a cluster.
- Library-embedded. mast composes as a Go library inside a larger service, not only as a standalone binary. Both shapes are first-class and share the governance layer — the agent loop, the dispatch shapes, the write gate, the effect outbox, budget ceilings and the watchdog are one implementation, and the session log is the same log. What the library does not carry is the plumbing that starts a turn with nobody calling: schedules, the monitoring cycle, notify, auto-resume and the operator listeners are the binary’s, on the assumption that a host service already has its own. The slim-embed guarantee is CI-enforced: you pay for what you import.
- Multi-provider. Two vendors — Gemini and Claude, each first-party or on Vertex — and the property worth buying is substitutability rather than the number. The same config switches between them without code changes, down to a single specialist; the meter prices the (backend, model) pair; and a judged corpus runs nightly against both, so “the same workload behaves over there” is a measurement. Model quality moves month to month, and being pinned to one vendor’s curve is a risk a platform team should not have to carry.
- Durable. Sessions survive process restarts, pod rescheduling, and
kill -9. A paused approval outlives the process that asked the question. Unattended without durable is unwatched but fragile — which is why durability is a pillar and not a storage detail.
Where mast fits
Section titled “Where mast fits”Two honest routings — the sooner you get one, the better:
| If you are… | Use | Why |
|---|---|---|
| Someone who needs one simple agent loop in Go | raw ADK v2 | Smallest surface, no governance, no durability. That niche is ADK’s, and wedging a thin mast-branded layer in front of it would shrink every time ADK improves. |
| A platform or SRE team putting agents where no human is watching | mast | Keep reading. |
Come back from ADK when the loop must survive restarts, needs budget or permission governance, needs provider switching, or stops having a human watching it. Adding those is an import, not a migration — you start inside mast either way.
The scope decision behind that: mast invests in the deployed shape — durability, governance, operator surfaces, provider portability — and not in interactive editing UX. There is no LSP integration, no AST tooling, no syntax-aware diff UI, and the file-editing tools exist to serve unattended workloads rather than an editor. That is a boundary chosen once, not a backlog item.
Where mast wins
Section titled “Where mast wins”- An agent runs inside a Cloud Run pod, a Kubernetes operator, or a scheduled job, mostly unattended.
- An agent is a Go library compiled into a larger service.
- The same config has to switch between Gemini and Claude.
- Multi-tenant auth, audit logs, cost ceilings, and permission gates are first-class requirements rather than someday-items.
- The workload is platform- or SRE-shaped: incident triage, deployment automation, drift detection, runbook automation.
Where it does not
Section titled “Where it does not”- The agent’s home is a developer’s editor rather than your infrastructure.
- The work is open-ended exploratory bug hunting in unfamiliar codebases.
- LSP-style symbol and AST awareness is core to the workflow.
- One process, one loop, one human watching — no durability or governance needed.
The sibling, not the successor — and now neither
Section titled “The sibling, not the successor — and now neither”mast is a lean fork of core-agent, and both are maintained. They have different jobs: core-agent is the broader experimentation and integration substrate, with an interactive TUI and a wider surface; mast is the narrow, unattended, deploy-shaped one. Neither replaces the other, and a decision that assumes “mast is the future core-agent” is starting from a false premise.
The full argument, including the kept/cut/change-shape decisions this page
summarizes, is in
docs/positioning.md.
- Concepts — how the pieces fit together.
- Quickstart: unattended triage — the whole loop offline, no credentials, in about five minutes.
- Roadmap — what is shipped, and what is honestly not.