DOP
GitHub
Menu
Decisions ADR-0007 2026-08-29

Threads you can address, findings you can publish

A demand's conversation is a set of threads — the main agent and one per specialist — each with its own history and card. Specialists share what they found as structured findings, not by copying their whole context into each other.

What was on the table

A demand may need several agents at once: one implementing, one reading a database forensically, one combing logs. The developer needs to talk to each without mixing their timelines, and each needs to use what the others found. Inside a demand the agents share a sandbox and cooperate; between demands the boundary is hard.

The paths we weighed

A single sequential agent. Rejected: no specialisation, no parallelism.

One sandbox per subagent. Rejected: it breaks the shared workspace the specialists need, multiplies cost, and isolates agents that cooperate.

One shared timeline. Rejected: separate timelines are the requirement.

What we chose, and why

The demand’s conversation is a set of threads: #main plus one per subagent, each with its own history; the developer addresses one thread at a time. Every subagent has a card — purpose, tools granted, the model the router chose, a slice of the demand’s budget.

Cross-thread knowledge is by query, not by dump. Threads are readable by their siblings through tools; no timeline is copied into another’s context. A conclusion is a published finding — a structured result on the demand’s board that enters the siblings’ context, is an event, and is written to the project’s memory.

Subagents are launched by the human, through the chat, or by the main agent on its own initiative — an assumption the product may restrict. The security boundary stays the demand: subagents share the sandbox, the workspace, the credential and the quota.

What it cost

The runtime supports several sessions per sandbox, and the attention box becomes a prerequisite for scale rather than an option.

Since then

The quarantine of raw tool output in the specialist’s thread is also the largest saving in the cost decision: the main agent receives the finding, never the dump.

Decisions this one leans on

  • ADR-0004
    Every write is an event; everything else is a projectionEvery write in the core emits an immutable event carrying who, what, when and in which context. The dossier, the timeline, auditing, replay and metrics are read from that log — none of them is written separately.
  • ADR-0006
    Three layers of knowledge, one curated package per demandA project's knowledge is rules, an index of its code and a memory of past demands. Each demand receives a curated package assembled from them, and the whole shelf is cloned into the sandbox for the agent to open what it needs.
  • ADR-0008
    Measure, cap, route — and design the prompt for the cacheEvery model call emits a cost event; every demand has a budget that pauses rather than kills; a router picks the class of model per kind of work — never cheaper for the critic. And the prompt is laid out so the expensive part is cached.
  • ADR-0017
    One microVM per demand, one worktree shared by its threadsThe hard boundary is between accounts and between demands, so each demand gets one sandbox and its agents share one workspace. Verification never runs there: it runs in a separate environment built from a commit.