Six ideas you won't find assembled anywhere else.

Most of GoTo Agents is deliberately boring — files, folders, open standards. These six ideas are where it departs from the rest of the field. Each gets a paragraph here and a deeper page of its own.

1. Agent-Initiated Requests (AIR)

Everywhere else, when an agent lacks a capability, it fails or hallucinates around the gap. Here, the agent asks the calling side — you, or a connected client like Claude Code — with a structured request fulfilled through conversation, not RPC. The other side can answer, clarify, partially fulfil, or decline, and the agent adapts. Humans and callers stay first-class participants, not execution backends. Deep dive →

2. MCP Exposures: capabilities as declarations

A capability can exist before its implementation does. Declare a proxy exposure (tools declared inline, every call routes to the caller) or a pure capability stub (an interface with no implementation at all), and agents can plan around it, attempt it, and exercise it via AIR. Prototype a tool by playing its part in conversation; implement it later without changing the declaration. Deep dive →

3. Activities: modes, not just prompts

Real collaboration has modes — brainstorming is not drafting is not reviewing. An Activity is a playbook-driven working mode you and the agent enter together, mid-conversation, and leave when the work changes shape. It reshapes the whole conversation toward a kind of output, not just one reply. Deep dive →

4. Memory built for days, owned by the work

Agent memory elsewhere is per-assistant and hidden. Here, memory belongs to the work: one shared MEMORY.md per project that every agent reads and writes, plus one per profile for what agents learn about you. Both are plain markdown, one click from view, one edit from corrected — and because project memory lives in the project folder, copying the folder carries the shared understanding with it. Deep dive →

5. MCP in both directions

Attach MCP servers to give agents capabilities — standard. Less standard: GoTo Agents is itself an MCP server, so Claude Code, Claude Desktop, or any MCP client can drive it — run tasks, message agents, read definitions. Combined with AIR, the relationship becomes genuinely bidirectional: the client that drives GTA can also lend GTA agents its own capabilities.

6. Open formats all the way down

Skills follow the cross-vendor Agent Skills standard (SKILL.md) — ecosystem skills drop in unchanged, and yours work everywhere else. Project instructions follow the AGENTS.md convention. Memory follows the conventions modern models are trained on. Every object in the system is markdown-with-frontmatter you can read, diff, and version. If you ever leave, everything comes with you.

What we deliberately didn't build

No swarm orchestration, no autonomy loops, no agent-to-agent protocol, no embedded vector database, no framework DSL. For fully-specifiable tasks, automation tools do the job — use them. GoTo Agents is for the work you can't fully specify up front, because discovering what you want is part of the work. That's not a missing feature; it's the thesis.