Colleagues, Not Tools

v1.0 made a decision that looks backwards: not giving the main agent more tools, but narrowing its tool surface. An all-powerful agent that does everything itself is a black box nobody can audit. So Previously is a team: the main agent owns the conversation and the judgment, and specialist work goes to specialist colleagues.

The main agent's six tools

Six tools remain, each with an explicit reason to exist:

• recall — ask the past (see Recall).
• webSearch — ask the world (see The Web Researcher).
• thinkDeep — ask a mind the conversation never touched (see The Clean Room).
• readSlice — the one memory-reading tool it kept, reserved as a verification-only channel: open the raw slice only to audit what recall reported.
• readPreviously — read the user card and the current belief snapshot.
• currentTime — the agent's watch check. Precise time must come from a tool, because the system prompt is byte-frozen within each slice — the precondition for provider prefix caching.

Every memory-browsing tool — readTimelineWindow, readStrand, readSliceSummary, and the rest — moved down into the colleagues. The main agent no longer walks the archive itself. And every tool is conceptual: it sees slices, strands, and timelines, never filesystem paths.

Narrow is auditable

A narrow surface isn't a capability cut — it's where auditability comes from. Everything the main agent can do that touches memory or the outside world is a named call with a payload and a report, reviewable one by one after the fact. A god-agent's "I looked around and felt that..." can't be checked; a recall call's question and report are on the record.

Trust, but verify

Trusting a colleague doesn't mean giving up the audit — and the audit is structural. When the main agent re-reads a slice after a recall, the read is classified and scored: a quick check is a verify, re-doing recall's job is a rework, asking the same question again is a repeat. A rework is an implicit demerit against the colleague, and enough of them trigger an evolution (see The Evolution Loop). Not "we trust our colleagues" — "the system keeps score of them, continuously."

One runner, one set of house rules

All colleagues run through one shared runner with a shared base prompt that pins the relationship: the caller is a colleague, and the user is always a third party — nobody role-plays you, and nobody treats the other side as you. Colleagues stream live, time out gracefully with partial results, and never throw an exception into the chat.

The memory gatekeeper

Every turn also runs one structured housekeeping pass — the turn analyzer. It doubles as the memory gatekeeper: its memory_worthy verdict decides whether the turn earns a place in memory. Small talk and content-free turns stop here. Not every sentence deserves remembering — trustworthy memory starts with refusal.

Related links

• Recall — the first colleague: evidence anchoring and the rework signal
• The Web Researcher — the researcher colleague
• The Clean Room — the isolated-reasoning colleague
• The Evolution Loop — how the scorekeeping becomes improvement