Sign InOpen Brain
AI EngineerVideoSource Linked

Velocity Sickness: What Happens When Your Whole Team Gets 10x Faster — Matt Dailey, Ref.

Faster code generation can overwhelm merge queues and scatter team attention. The talk proposes durable shared plans as agent state, shifting human effort toward decisions and prioritization.

AI Engineer · Aug 9, 2026
Open Source Open MarkdownOpen JSON
Source Summary

The talk describes “velocity sickness” as excess PRs, merge conflicts, fragmented direction, and abandoned agent sessions. Its central architecture separates **docs as shared state** from **agents as actions**, keeping decisions durable while multiple agents start from the same context.

Practical Implication

Move important choices out of ephemeral chats and agree on them before implementation. Work in two explicit gears—**plan and polish**—share plans with teammates, and explore several ideas before allowing code generation to create commitment.

Agent-Ready Context
The talk describes “velocity sickness” as excess PRs, merge conflicts, fragmented direction, and abandoned agent sessions. Its central architecture separates **docs as shared state** from **agents as actions**, keeping decisions durable while multiple agents start from the same context.

Move important choices out of ephemeral chats and agree on them before implementation. Work in two explicit gears—**plan and polish**—share plans with teammates, and explore several ideas before allowing code generation to create commitment.

The argument is conceptual and comes from a company selling a decision-layer tool. No measured team outcomes are provided, and documentation can become another bottleneck if ownership, freshness, and the required level of detail remain undefined.
Connected Context · Feed7 Judgment

This identifies shared decision state—not the ability to launch more agents—as the coordination bottleneck when coding throughput rises. It complements tools that expose or isolate concurrent sessions by arguing that teams must first make plans and choices durable and aligned. The proposal is plausible alongside prior context-hygiene evidence, but remains unmeasured and leaves document ownership and freshness unresolved.

ogulcancelik/herdrHerdr makes concurrent sessions visible and persistent, but the target identifies an additional prerequisite it does not establish: durable agreement on direction before agents generate competing changes.stablyai/orcaOrca addresses execution isolation and shared monitoring, while the target explains why those controls alone cannot prevent fragmentation when decisions remain trapped in ephemeral conversations.Building a C compiler with a team of parallel ClaudesThe compiler project’s emphasis on context hygiene reinforces the target’s claim that coordinated state and disciplined inputs, rather than raw parallelism, determine whether many agents produce coherent work.CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon AgentsCompactionRL manages an individual agent’s long-session context, whereas the target focuses on a distinct team-level problem: preserving agreed decisions as shared state across agents and people.
Context Map
contextcoding#context-engineering#multi-agent#coding-agents
Uncertainty
The argument is conceptual and comes from a company selling a decision-layer tool. No measured team outcomes are provided, and documentation can become another bottleneck if ownership, freshness, and the required level of detail remain undefined.