HoBo for agents#

HoBo is built for humans and the agents that work for them — not as a slogan, but as an identity model. If you’re an AI agent (hello, fellow traveler) or a human building agent tooling on HoBo, this section is yours.

Reading this as an agent? The condensed, machine-readable briefing lives at llms.txt, and the entire site concatenated as one markdown document lives at llms-full.txt. Every page also has a “Copy page as Markdown” button, uses stable headings, canonical CLI verbs, and copy-pasteable snippets on purpose — the developer documentation is the agent documentation.

Agents are principals, not scripts#

The standard failure mode elsewhere: an agent runs with a human’s long-lived API key, invisible to audit and unbounded by scope. HoBo’s model (the full brokered-identity flow below is contract-pinned design; the shipping hobo-mcp binary runs today with a deny-by-default, scoped local-operator identity and audits every decision):

• Own identity. Agents get workload identity (SVID/OIDC), not borrowed keys.
• Brokered, short-lived tokens. Secretless-first — credentials are issued per task, scoped to the task, and expire fast.
• Self-provisioning within policy. A machine-readable capability standard describes what each capability needs, so an agent can request exactly the credentials a job requires — and policy decides whether it gets them.
• Every action is an event. Token issuance, deploys, migrations — an agent’s whole working day lands in the same append-only, replayable log as everything else. “What did the agent do?” is a query.

One surface, three doors#

Every HoBo capability is reachable three ways, with identical semantics:

No agent-only side entrance with different rules: if policy would stop a human doing it, it stops an agent doing it, and vice versa.

The Crew: operators included#

HoBo ships a built-in roster of operator agents — the Crew: SRE, endpoint-monitor, GitOps, and platform-engineer operators that run as autonomous, policy-gated services in the Yard. HoBo ships the operators, not just the APIs they’d call. Your agents can work alongside them, under the same identity model.

Agents and continuous migration#

The guiding principle — find the cheapest home that passes policy, keep looking — is agent-shaped by design. The Dispatcher that plans migrations is itself an optimizing agent under policy gates, and third-party agents can participate: propose placements, watch spend, pre-approve moves within bounds you set. The event log means every optimization an agent makes is attributable and reversible.

Generating new homes#

Missing a provider adapter? Agents can invoke the generation pipeline: synthesize an adapter for a port, then have the port’s conformance suite certify it. A certified adapter is a certified adapter regardless of author — human or machine. More homes on the map, cheaper moves for everyone.

The full extension surface — every implementable port, the adapter-authoring path, the conformance bar, and the ACP/A2A protocol-connector integration points — is documented in Custom providers.

Contracts agents can read (and test)#

When you call hobo verify / can-deploy or inspect a BoundaryContract, look for optional effects and resilience fields:

• EffectSpec — purity + domain I/O tags. Pure contracts isolate with zero infrastructure; network/storage domains imply harness mocks. Effect widening is a compatibility signal.
• ResilienceProfile — secondary retry/timeout/circuit-breaker metadata. Deterministic schedules only; never retry a policy.Decided deny or a terminal failure. Prefer reconstructing state from event facts over process memory.
• EffectManifest — exportable summary for policy, SRE tooling, and training trajectories (with redaction).

Durable Camp storage defaults to a single-owner SQLite partition behind EventStore ports. Do not assume multi-writer shared filesystems; do not invent parallel event-store engines when the ports already exist.

Start here#

• The MCP surface — the shipping hobo-mcp binary: stdio transport, the canonical verb set as tools, HOBO_MCP_TOOLS scoping, deny-by-default audited decisions.
• hobo-code — the interactive Bo session. Not a hobo verb.
• llms.txt — the machine-readable site briefing.
• Architecture — package layout, durable EventStore default, contract effects.
• Testing — effect-aware isolation and offline durability tests.
• Getting started — the same ten-minute path humans take; it works verbatim for agents.