An agent watched it happen.

Ask the agent, “What were they thinking?”

When your teammate finishes a hard piece of work with an agent — a PR, a pricing investigation, a research synthesis — you don't have to ask them what they were thinking. The agent that watched it happen is right there.

  • “Why did she rule out polling?”
  • “Why did we kill the free tier?”
  • “What did the customer interviews actually rule out?”

The agent answers as the stranger who spent the afternoon with them. Sapwood is the substrate that makes it possible — and the trail of decisions that builds up across sessions, teammates, and time.

Works with the agents your team already uses. Lives in the tools you already work in.

See how it works

The problem

Your agents got faster.
Reasoning didn't keep up.

Three pains show up wherever knowledge work meets agents.

The vanishing. A teammate (or you, last Tuesday) spends three hours with an agent untangling a hard problem — sketching options, ruling out dead ends, landing on the right framing. A week later, no one can reach it. The conversation is buried in a chat history that's unsearchable, unbranchable, useless for continuing the work. The reasoning is intact and inaccessible.

The summary trap. The work has to leave the session somehow. A Slack message. A standup. A doc. What comes out is a rough summary — three hours of reasoning collapsed into a few bullet points. The reader builds a plausible-sounding mental model from the summary and acts on it.

The parts of the reasoning that got dropped on the floor are exactly the parts that come back to bite you three weeks later.

The bottleneck. Every handoff waits on the person who did the work to re-explain it. Every re-explain leaves more out. Your organization's rate of understanding is capped at whatever rate one person can narrate their own brain — to a colleague today, or to themselves next quarter.

These compound. The work doesn't accumulate. It doesn't transfer. It doesn't scale.

This is the new bottleneck of knowledge work — not building, but keeping the reasoning intact, accessible, and shared.

How it works

One surface for the work, the questions, and the record.

Four primitives. One idea: the work itself becomes the record.

01

Save and resume

When you finish a productive stretch with an agent, save the session. Anyone with the URL can resume it: pick up exactly where it left off, in a fresh session that inherits the full conversation. Not a summary of it. The actual tokens, the actual context, the same agent that watched the work happen. We call these ambers — and warming one is how you pick it up.

02

Pins

A lightweight anchored marker on a moment in a session — a question, a note to self, a “come back to this.” Pins can be private, shared with your tree, or assigned to a teammate. Optional sidebar conversation. They never block the work.

03

Decisions, on the record

Some decisions need to mechanically pause the work — webhooks or polling, build or buy, ship now or wait. The session pauses at the question. Engagement is the unblock. When the discussion resolves, it produces a verdict — a first-class decision record that stays attached to the work that shaped it. Substantive, abandoned, deferred — every decision resolves to something. We call these burls.

04

Branches

Three kinds, all preserving the trajectory the original session built up. Branch to extend a line of thought (continuation). Branch to interrogate the prior author's thinking (more on that next). Or just open a fresh session if you need to escape a stuck framing entirely.

05

Lives where you already work

Sapwood meets your team in the tools they already have — not another app to migrate to.

  • GitHub A comment on every PR with the agent session that produced it. Reviewers click “Resume in Claude Code” right on the PR.
  • Your terminal A skill that ambers what you're doing without being asked. Session-end hooks capture every productive stretch into your team's tree, automatically.
  • Your browser A Mac app and a browser extension hand off cleanly: any sapwood:// link opens a warmed session in your terminal, or in the treehouse, or wherever you want it. Coming soon
  • Slack, Linear, Notion Paste a warm command anywhere to land in a session in one click. Coming soon

How this is possible

The agent already knows how they think.

The agent in your teammate's session didn't just produce output. It spent two hours being your teammate's interlocutor — modeling their priorities, their vocabulary, their reasoning style, what they'd push back on, what they'd reject. That cognitive work shaped every response.

When you warm the session, you inherit that work.

A summary can't do this. A doc can't do this. A recording of a meeting can't do this. The substrate has to be the conversation itself, preserved at the token level — and that's what Sapwood is.

The same mechanism serves future you. Past-you's session is a stranger's session. The agent that was there can tell you what past-you was thinking, in conversation.

Who it's for

Two axes. The same bottleneck.

Sapwood is built around one idea: the reasoning is the work. That idea shows up on two axes — across people and across time — and Sapwood treats them as the same problem.

Working alone, over time

Pick up where past-you left off.

The PM who spent three hours mapping competitors last Tuesday. The engineer who untangled a billing edge case in a session they can't find anymore. The researcher whose best thinking lives in a chat history that doesn't survive next month.

  • Engineering Warm last week's debugging session into a fresh agent that knows the whole story
  • Product Competitor research that compounds across launches, not lost between them
  • Both Interrogate past-you's session: “what was I actually thinking?”

Working together, in real time

Stop briefing. Start branching.

Engineering teams reviewing a PR that drops every alternative the agent ruled out. PMs and researchers pasting a warm command in a Slack thread, sending teammates straight into the pricing investigation — instead of flattening two weeks of strategy work into a deck. Skip the flattening step.

  • Engineering Every PR carries the agent session that produced it. Reviewers click “Resume in Claude Code” on the PR
  • Product A two-week pricing investigation lands as a warm command in Slack. Teammates branch into the session — no deck, no briefing
  • Both Decisions become searchable verdicts — attached to the work that shaped them, not lost in a channel

Trust

Your team's reasoning, kept where it belongs.

Sapwood captures your team's hardest thinking. We treat that seriously.

  • Encryption

    Encrypted at rest and in transit. Customer-managed keys (CMKs), cross-account S3, and AWS Clean Rooms are available on request.

  • Hosting

    SaaS on AWS with full tenant isolation. Self-hosted deployments on the roadmap for teams whose data residency requirements demand it. Self-hosted: coming soon

  • Capture controls

    The CLI filters out .env, secret-bearing paths, and other sensitive files by default. The blacklist is configurable. A pre-upload interview lets you review and approve exactly what gets included in any session before it's shared.

  • Access

    Fine-grained ACLs at the session level. Grant access by tree, by individual session, or to a specific reviewer. A PR comment stays visible to anyone who can see the PR; the session it links to opens only for the people you've granted.

  • Training

    Sapwood doesn't train models on your content. Your team's reasoning isn't a training corpus.

  • Audit

    Every amber, every burl, every verdict is logged. Trees produce a structured record of who decided what, when, and why — not as an afterthought, as a byproduct of the work.

Early access

Join the waitlist.

Sapwood is in private development. We're onboarding a small number of teams — and a small number of solo operators — whose work is already happening inside agent sessions. Tell us about yours.

Or email us
We read every response.