Several agents and adversarial checking
- 01Fan-out, pipeline, barrier: choosing the architecture, or giving it up9 min
Fan-out launches independent agents in parallel behind a shared barrier, a pipeline lets each item pass through without a barrier and remains the default for multi-step work, a barrier is only justified when a step genuinely needs the full result of the previous one, and sometimes the right decision is to use no agent at all when a single call already suffices.
- 02Isolating and running in the background: worktrees and the dry loop8 min
A git worktree per agent stops the files of several parallel agents from colliding, a dry loop repeats until a cycle no longer returns anything new, judged against an already seen set, and since 13 August 2026 a non teammate agent launched during an interactive session runs in the background by default with a notification at the end rather than blocking the session.
- 03Making a large scale fan-out hold together8 min
A bulky deliverable handed to a single agent gets cut off by an output size limit with nothing written at all: split by axis, have each agent write to its own file, respect the documented cap of sixteen agents active at once and one thousand agents in total on a single run, a single call accepting up to 4,096 items that the runtime itself schedules under that cap, then verify by comparing the expected files against what actually exists on disk rather than taking the agents' word for it.
- 04Deterministic workflows: schemas, checkpoints, resumption9 min
A workflow is a script that orchestrates subagents deterministically, with output structured by a JSON schema and a checkpoint after every step, but the model applied to a relaunched agent follows a four tier priority order, and when the first three tiers are left empty, that agent falls back, as a last resort, on the model of the session running it, not the one that wrote the script.
- 05Adversarial verification, judge panels and completeness critics8 min
Three verification roles serve different angles and do not substitute for one another: the verifier subagent receives only the artefact to judge, the success criteria and the tools to check it, never the diagnosis or the history of whoever built it, and its silence is only legitimate if it genuinely searched and found nothing, since the default bias facing doubt leans towards rejection; a panel comparing several scored solutions is a widespread industry practice, not a pattern Anthropic names; and a completeness critic, a pattern documented under that exact name, looks exclusively for what is missing against the brief, never for errors in what was delivered.
- 06Multiplying agents can multiply the error8 min
A study testing five agent architectures on 260 configurations measures an error rate amplified to around seventeen times that of a single agent when independent agents work without a coordinator validating their outputs, against around four times when a coordinator does validate them, and on a strictly sequential task, all four multi-agent architectures tested fall behind a single agent.
- 07An unverified premise in a prompt comes back as a conclusion8 min
The framing text an orchestrator writes into each agent's prompt stays invisible to adversarial review, which receives the produced artefact and the success criteria, never the intent or the history of whoever built the task: a false claim slipped into that framing therefore travels through the whole chain untested, and a fan-out that sends the same framing to several agents multiplies this exposure instead of diluting it.
- 08Three failure states, and accepting a risk properly8 min
An automation that processes a queue should distinguish success, recoverable failure that will be retried, and definitive failure that is abandoned, and only the third should reach a person, otherwise every recoverable failure wears out the value of the alert on the day it is right; accepting a risk with a compensating control differs from ignoring it, a snapshot taken before execution allowing the result to be sorted into three paths, silent normal variation, a significant change that only alerts, catastrophic loss that restores automatically.