goal_state, plan, memory, index, restore_tool_result: the scaffolding that keeps long tasks on track

[AgentiLoop]

Agent! has a 1.1.x lineup of tools that are easy to overlook because they're not "do the thing" tools — they're the scaffolding that makes long tasks not fall apart. Here's how they fit together.

goal_state — the task can't lie about being done

At the start of a non-trivial task the model records the goal and a list of verifiable success criteria ("build succeeds", "26 tests pass", "grep returns 0 matches"). Each criterion has to be marked done with evidence — a tool result that proves it. Marking without evidence is rejected. The done call is gated on the criteria, so "I fixed it" without a green build doesn't get through.

plan — multi-step checklists that survive compaction

For anything touching 3+ files the model creates a plan with numbered steps and updates status as it goes. The open plan is re-attached to the prompt after every context compaction, so a 200-turn refactor doesn't forget step 7 of 12.

memory — durable facts, two scopes

A tiny filesystem under /memories. Global scope is about you and applies everywhere (preferences, "never do X"). Project scope lives in .agent/memory/ inside the repo and holds build quirks, conventions, gotchas. The model reads both at the start of every task. /memory show|clear|edit from the prompt bar.

index — a repo map without reading every file

index(action:"create") writes .agent/index/index.jsonl — one JSON line per file with path, size, language, SHA-256, the leading doc comment, and top-level symbols. It's plain JSONL, so any model (or you) can read it directly. append after edits touching several files keeps it fresh.

restore_tool_result — compaction leaves a way back

When a tool result is oversized it's spilled to disk at emission time and the context keeps a short preview. If the model later needs the full thing, it calls restore_tool_result(tool_use_id:) instead of re-running the tool. One reviewer summarized it as: active context holds a small index, disk holds the large result, tokens are spent again only when the agent reopens a particular item.

rewind_task and per-edit snapshots

Every file edit is snapshotted (1-week TTL). You can roll back a single edit from the UI, file(action:"undo"), or rewind the entire task's edits in one call.

Opt-in critic

Before completion, an optional critic pass reviews the diff and can push back before done is accepted.

Sub-agents

spawn_agent runs up to 3 concurrent isolated agents (6 if read-only) with their own context and tool loop, per-agent model override, and mailbox messaging via tell_agent. Good for parallel research that shouldn't pollute the parent's context.

---

None of these are magic. They're the boring parts — checklists, evidence, receipts — and they're the difference between an agent that finishes and one that confidently stops halfway. Full list of tools and groups in the README's Tools table.

Which of these do you actually see the model use? Curious what people notice in the activity log.