Slack engineers published an architecture breakdown — covered by InfoQ on April 28 — describing how they manage context in long-running multi-agent systems where one application can span hundreds of requests and generate megabytes of output. Standard agent frameworks accumulate chat history between calls, but in long-running production sessions that approach hits the context-window ceiling and degrades response quality well before the hard limit. Staff engineer Dominic Marks describes the alternative: replace accumulating chat logs with structured memory channels, dedicated critic agents, and a distilled timeline that survives across the lifetime of the run.
The architecture is a coordinator/dispatcher pattern. A central coordinator receives requests and routes them to specialized agents downstream — experts that do the actual work, and critics that evaluate the experts' output. Critics are necessary because (in Slack's words) a portion of expert findings "could either be invented or grossly misinterpret the data." Three structured context channels carry the running state instead of an unbounded chat log. The director's journal stores the director's structured working memory — findings, observations, decisions, open questions, hypotheses — and is described as "the common narrative that keeps other agents on track." The critic's review is an annotated findings report with credibility scores, built by critics narrowly instructed to "only make a judgement on the submitted findings" — that is, don't drift, don't invent. The critic's timeline builds a coherent narrative from the director's journal, the latest review, and the previous timeline, deduplicating findings and resolving conflicts by preferring credibility-weighted evidence.
The pattern matters because the chat-log accumulation strategy that works for short single-agent interactions does not scale to multi-step production workflows. Three things Slack's design implies are now industry common sense for long-running agents: critic agents are necessary because expert agents hallucinate at meaningful rates; structured memory beats raw chat history because it forces summarization and credibility scoring; and time-ordered narrative-distillation — Slack's "timeline" — is needed to keep agents coherent across hundreds of requests. The pattern is generalizable: any team running multi-agent workflows that span minutes-to-hours instead of seconds will need some form of director/critic/timeline split, or a close equivalent. The exact channel names and structures will vary, but the three roles — narrative carrier, credibility scorer, timeline builder — are likely to converge across frameworks.
For builders, three concrete things. First, if you are building agentic workflows that span more than a single chat-window's worth of state, do not just accumulate message history — design structured memory channels with credibility metadata from day one. Retrofitting that pattern later is painful. Second, critic agents constrained to "only judge submitted findings" are the cheapest hallucination defense Slack found: narrow scope plus dedicated role plus credibility scoring. Build that into your multi-agent layer rather than hoping the expert agent will catch its own mistakes. Third, the engineering culture matters: Slack is publishing this architecture publicly, which means future open-source frameworks will likely standardize on something like coordinator/director/critic/timeline. Watch LangGraph, AutoGen, and CrewAI for this pattern landing as a first-class abstraction over the next several months.