cross-provider agent swarms
Test a shared instruction across providers, give planning and review to a strong agent, distribute repeated implementation to cheaper workers, and keep the handoffs inside one runner.
Show Us Your Agent Skills / EP 08 / guest dossier
CHIP HUYEN
AI ENGINEERING
MULTI-AGENT SYSTEMS
GOOD AI STACK
CHIP HUYEN
Chip showed a Good AI Stack project containing 1,409 agents, 1,951 instructions, and 330 recorded lessons. Her runner puts a strong model such as Fable in charge of planning and review, then hands repeated implementation to cheaper workers across providers. For each task, it records what Chip asked, which agent handled it, whether it finished, and the evidence the agent returned.
EP 08 · CHIP HUYEN · cross-provider agents, task memory, skills, and failure lessons
THE RUNNER HANDLES THE HANDOFFS
a strong agent plans and reviews; cheaper agents implement across providers
Chip used to copy plans and review comments between model-provider tabs. Her runner lets one agent ask another to review its plan, receive the response, and assign follow-up work in the same conversation.
The supervising agent spends its reasoning on the plan, worker selection, review, and deciding whether the assigned task is finished. Once Chip is happy with the guideline or plan, it can send repeated work to Sonnet, GPT, DeepSeek, Kimi, or other cheaper agents.
The Good AI Stack project she showed contained 1,409 agent entries, 14 skills, 1,951 instructions, 33 known issues, and 330 lessons. Chip can return to the runner instead of reconstructing the job from open tabs.
"I get a very strong agent to make plans and think, and then it assigns tasks to a bunch of cheaper agents to do it."
The strong model stays on the decisions that affect every worker. The repeated implementation goes to faster, cheaper agents. 00:50:08
TEST THE GUIDELINE, THEN MULTIPLY
the Good AI Stack taxonomy job, from one shared instruction to a large labeling run
Choose one bounded exampleGive the same product and taxonomy guideline to agents from different providers. 00:46:35
Compare their labelsDifferent providers expose weak instructions that agents from the same model family may interpret in the same wrong way. 00:47:38
Revise where they disagreeChip updates the guideline and repeats the small comparison until the category boundaries hold up. 00:47:48
Give the plan to the supervisorThe strong agent keeps what Chip asked in view, chooses workers, and reviews their output. 00:50:08
Dispatch the repeated workOnce the guideline holds up, the supervisor creates cheaper workers across providers and keeps them moving until their assigned work is complete. 00:49:46
Collect and review in the runnerResults return to the supervising agent without Chip carrying messages between tabs. 00:54:28
"I care more about what I think of as human context management."
Agents can work across many tasks. Chip can attend to two or three, then a meeting or another project breaks the thread. 00:56:37
BUILD THE HUMAN'S MEMORY INTO THE RUNNER
what Chip asked, which agent handled it, its status, its evidence, and the failures worth remembering
Every instruction can become a task record. The runner preserves what Chip asked for, which agent received it, whether the work finished, and the evidence that agent supplied. When a bug returns, she can inspect the record instead of searching every conversation.
What Chip asked stays visible during planning and implementation. Otherwise, an early misunderstanding can become the plan, then the plan can replace the task she actually gave the agent.
Failures become records too. Chip classifies planning errors, execution errors, instruction-following failures, and jobs that should have used a deterministic script. That history helps her decide which model should receive similar work next time.
"It's horrible for my mental health."
Chip's answer to what she loves about working with agents began with the mistakes. They have made her more patient, and shown her where her own instructions were weak. 01:16:23
MAKE THE AGENT SHOW THE PAGE
Chip's verified frontend skill turns a rendered browser view into review evidence
When an agent changes an interface, Chip's verified frontend skill tells it to open the site in a browser, inspect the rendered result, improve visible problems, and capture a screenshot.
The screenshot travels with the pull request, so Chip can review the visible result without checking out and running every branch. The browser mechanism depends on the environment: the desktop app can use its built-in browser, while a terminal or CI run needs a local headless browser.
Chip wanted a script to detect that environment. Asking another agent to reason about a deterministic choice would add another instruction that could fail silently.
CHIP'S OPERATING KIT
the workflow and skill reconstructed from what she showed
verified-frontend
Open frontend changes in a browser, inspect the visible page, fix what is visibly wrong, and return screenshot evidence with the work.
taxonomy labeling
Define category boundaries, apply them to the same products through different providers, and revise the guideline where their labels diverge.
write tests
Stop agents from turning trivial copy changes, removed behavior, or the wording of a skill file into permanent regression tests.
THE STACK
the tools, models, and boundaries around Chip's runner
Codexcoding-agent product used for planning, implementation, and review
Claude Codestrong-agent planning and mobile handoff
DeepSeekcheap open-weight implementation workers
Kimianother lower-cost worker option
Git worktreesisolation for every code-changing agent
GitHubrepository data and screenshot-backed review