For teams building with coding agents

Agents do the work.
You stay in command.

Maalstrom runs product work from the first scrap of an idea to shipped code. Agents research, design, plan and build. People step in where a decision needs judgement, and everything that happens along the way is on record.

Request early access See how it works →

The problem

We're the bottleneck now.

Agents can build faster than anyone can keep track of what they're building. The instructions they work from are scattered across chat sessions, documents, tickets and meetings, and keeping them right falls to people. Every person you add brings another set of silos.

How it works

Agents should execute,
not improvise.

Implementation should be deterministic, not probabilistic. Every stage settles something the next one would otherwise have to guess. Research narrows the question, design fixes the approach, and the plan fixes the steps, so by the time your agent builds, there's nothing left to improvise. Agents do the work inside each stage and bring you in at the boundaries, where judgement matters.

  1. 01CaptureNotes, links, conversations, half-thoughts. Nothing to fill in.
  2. 02IdeateRelated scraps are connected and worked into an idea with you.you shape
  3. 03ResearchAgents investigate. You make sure they're asking the right questions.you steer
  4. 04DesignAn approach is drafted and refined against your feedback.you review
  5. 05PlanConcrete steps an agent can carry out without guessing.you sign off
  6. 06BuildYour coding agent does the work, with exactly the context it needs.you oversee
  7. 07ShipThe changelog and documentation are written as part of the work.
  8. 08LearnWhat each build taught you shapes the next one.

Ideas

Start from a scrap, not a spec.

Nobody has the whole picture at the start. Drop in what you have, from wherever it comes from, and agents help turn it into something clear enough to act on. Nothing gets lost, and one scrap can feed several ideas.

  • Scraps come from you, your team, your meetings, or an agent mid-build.
  • Agents notice what belongs together and suggest where it leads.
  • You decide which ideas are worth pursuing.
  • Note · J. de Groot

    Nobody read past page 3 of the vendor whitepaper.

  • Link · M. Visser

    Summarising long documents with language models

  • Review comment · S. Bakker

    Which parts of this RFC actually matter here?

↓ 3 scraps, 1 idea

Asset summaries

Attach a long document, get the parts that matter to this piece of work.

worth doing?

In the document

Written together.
Agreed before it's built.

Every stage produces a document your team and your agents work on together. Shape it, comment on it, branch it and merge it back, then publish a version for the people it affects to sign off. Every version, comment and signature is kept.

3f9a2c1 Summaries per asset version 1d ago

Asset summaries

Attach a long document to an item, and get back the parts of it that matter to that item, without anyone reading past page three.

Context

Reviewers keep attaching whitepapers and RFCs to designs, and nobody reads them. Research on MAA-12 showed that ranking an asset's sections against the item finds the relevant ones reliably, where summarising the whole asset first loses them.

Storage

Summaries are generated when an asset is attached, and stored with the version that attached it. They are then regenerated again whenever somebody opens the asset after another author changed it.

Generation

Summaries are generated on request.

Rollout

Ship behind a flag, starting with our own org.

Highlighting

A summary picks out the sections of the asset that matter to the item it's attached to, ranked against the item's description. Each highlight links back to the passage it came from, so a reader can check it in context.

Data model

fieldtypenotes
asset_versionuuidthe version the summary was made from
itemuuidwhat the highlights were ranked against
sectionsjsonbranked highlights, each with its source offsets
modeltextwhich model wrote it, for later comparison

Risks

  • Scanned PDFs have no text layer. They need extraction first, which is MAA-16.
  • Ranking against a vague item description gives vague highlights. The pre-check flags items without one.
  • Summaries cost tokens on every attach. A shared cache per asset version keeps re-attaches free.

Alternatives considered

Summarise the whole asset first, then rank. Rejected in research: the summary loses exactly the details a reviewer is looking for.

Generate on first view. Cheaper for assets nobody opens, but leaves a reviewer waiting the first time, and an unopened asset without a summary at all.

Decision log

decisionbywhen
Store summaries with the asset version, not the assetproduct, engineeringv2
Rank sections against the item, not a fixed queryengineeringv3
Roll out behind a flag, our own org firsterikv4
⎇ mainMAA-14v4 published no open branches

Process

All of the process.
None of the overhead.

Specs, decisions, changelogs and documentation come out of the work itself, not out of extra hours. A founder working alone gets the record and standards of a much larger company from day one. As people and agents join, they start from all of it, and you add roles and sign-offs without changing how the work flows.

Coherence

Plans that stay true while everything changes.

Changes cascade. A shift in scope touches designs, plans, tickets and schedules, and in most teams nobody has time to chase every one. In Maalstrom the work is connected, so a change reaches everything it affects, and anything that needs a decision lands with the person who makes it.

For leaders

Know what's happening without asking.

Today, what reaches leadership is a report of a report, and the detail that explains a delay gets lost on the way up. In Maalstrom the record is the work itself, so the full picture is always there, from the overview down to the decision that moved the date.

IdeateResearchDesignPlanBuildShipLearn audit export awaiting sign-off summarizer agent building summary attachments plan changed comment anchors shipped offline drafts suggested

Fits your stack

Your agents, your tools, your infrastructure.

Maalstrom works with what you already run, and with your specs, plans and source, so where it runs is your decision too.

  • Claude
  • Codex
  • Gemini
  • Cursor
  • Mistral
  • Kimi
  • GLM
  • DeepSeek

Put your agents to work.
Stay in command.

Request early access Talk to us →