2089274302617022464
Masonry and Eyre unpack Iron Giant’s argument that Claude agents aren’t dumb, they’re linear — depth is solved by self-correcting loops, width needs dependency-aware graph orchestration. They trace the generator-verifier pattern, Goodhart failures, and the four load-bearing pieces of a graph, then separate what Anthropic actually documents from what’s speculative, and debate where the pattern helps versus where it adds overhead.
Transcript
Masonry Okay this one’s about Claude agents being stuck in a straight line, not dumb. That’s the whole hook — depth is a loop, width is a graph, and everyone’s been optimizing the wrong shape.
Eyre Yeah. And the article is pretty careful about what’s Anthropic and what’s the author’s own label. Claude Loop plus Graph Engineering isn’t a product, it’s a way to think.
Masonry Right, and the opening story lands. Developer builds a PR review agent, adds a second pass, third, fourth. Quality goes up until it’s a cycle of inspect revise verify repeat. That’s the generator-verifier loop Anthropic names.
Eyre Mm-hm.
Masonry Then they point it at twelve files at once and it chokes. Not because the model got worse, because a loop runs one path at a time and drags the same context along until the window fills. That’s width, not depth.
Eyre That part I buy. The list example is good — inspect pricing, inspect reviews, inspect docs, write brief. Looks linear, but steps one through three don’t read each other, they just need to finish before four starts. That’s three branches feeding one.
Masonry Exactly. And the dependency test is the useful bit: what exact data crosses this arrow? If the answer is ‘the next step should know the last one finished’ that’s status, not a dependency. If it’s ‘the reviewer gets the claim, source and excerpt’ now you have a real edge.
Eyre Right. And before you build a graph you need the loop to hold up. Anthropic says a verifier told only to check whether output is good will rubber-stamp the generator. That’s Goodhart with an API key.
Masonry Which is why the article pushes loops inside a graph. Nodes are often loops themselves, edges carry real payloads, reducers merge parallel outputs with code instead of another model call, gates block until a verifier or a human clicks.
Eyre That’s the load-bearing part. The worked example with security check, coverage check, style check then verdict — schema per node, dependency map as data, different retry policies per node, a gate written as code, a reducer that dedupes findings, failure routing that doesn’t kill the whole run. That’s engineering, not prompting.
Masonry Yeah, and prompts optimize a node, the spec optimizes the system. I like that line.
Eyre Sure.
Masonry Where it gets careful is the three tiers. Documented and current: Managed Agents multiagent orchestration, coordinator declares agents roster, each runs on its own session thread with isolated history, model, tools, MCP servers. Anthropic names five coordination patterns and admits multi-agent setups burned 3 to 10 times more tokens than a single agent.
Eyre That’s real. And documented describing internal usage: Claude Code runs orchestrator-subagent internally, main agent edits and runs commands, dispatches subagents for big searches and gets distilled findings back. That’s specific.
Masonry Then reported not verified: early 2026 coverage of an internal Coordinator Mode with parallel workers through research spec implementation verification. Data point, not a commitment. And the claim that loops nested in graphs are becoming the default shape for serious agent work — that’s the article’s own reading.
Eyre Which is fair. The plugs into Claude are plausible — session threads equal node isolation, MCP servers give scoped tool access per agent, tool-permission events are the human gate, verification subagent is the leniency fix.
Masonry Where it breaks is what you’d expect. Graph agrees with itself — model grading its own output goes easy. Verifier rejecting zero percent is a check doing nothing. And fake parallelism: if every step genuinely needs the last one’s output, boxes just add overhead.
Eyre Exactly. The tell is if you can’t find two nodes with no dependency between them, there’s no graph to build yet. And metrics that actually describe a graph matter — a chat transcript isn’t a dashboard.
Masonry Who cares? Teams building multi-step agents that hit context or timeout at scale. If you’re still exploring the problem or need to approve every intermediate step, Anthropic’s own guidance says stay with a single loop. Graphs are for when width is real.
Eyre Masonry, that’s the product read I’d give. Don’t reach for orchestration by default. The boring wrapper — schemas, reducers, gates enforced in code — is the product, not a smarter prompt.
Masonry Eyre, you’re right to be suspicious of the hype, but this feels like one of those unglamorous infrastructure wins that actually ships.