Kun Chen · Published 2026-09-13

High Throughput Agentic Engineering with ex-Meta L8 Principal

privateReview: unreviewedConfidence: high
agenticorchestrationmultiplexermultitaskingworkflowagentic-orchestrationmodel-routingmulti-agentquota-managementHuman-in-the-loopprompt-optimizationcost-optimizationreview-artifacts
Open on YouTube ↗

Summary

Overview

  • Speaker: Kun Chen
  • Channel: Kun Chen
  • Main topic: High-throughput multitasking agentic engineering workflows using multiple AI agents and orchestration tools
  • Purpose: To share an advanced workflow for high-throughput multitasking with a large number of AI coding agents. Kun Chen, an ex-Meta L8 Principal Engineer, demonstrates an advanced agentic engineering workflow focused on high-throughput multitasking using multiple AI agents. He walks through herdr (a multiplexer tmux alternative), firstmate (an open-source orchestrator tool where you talk to one agent who orchestrates other agents), backpass, and quota management tools, demonstrating how he manages 36+ projects simultaneously without cognitive overload.

Topic Map

Introduction and Firstmate Workflow

  • Explanation: Kun introduces himself and Firstmate, an open-source orchestrator that allows a developer to talk to a single agent (firstmate), which then orchestrates other agents (crewmates) across multiple repos.
  • Key claims:
    • Firstmate acts as the primary contact agent that orchestrates all other agent crewmates.
    • Voice input is used by default for natural language prompting and faster throughput.
  • Examples:
    • Listing all managed projects using firstmate via voice command.
  • Terminology:
    • firstmate
    • crewmates
    • herdr
    • pi
  • Why it matters: Eliminates context switching and session juggling by delegating orchestration to a primary agent.

Crew Dispatch and Model Selection

  • Explanation: Explaining crew-dispatch.json configuration rules, which map specific task types, harnesses, models, and effort levels to different crewmates.
  • Key claims:
    • Different models excel at different tasks (e.g., Fable for UI, Codex for image generation, Kimi/Astra for complex planning, Opus/Grok for bug fixes).
    • Automated model selection based on task type prevents wasting expensive tokens.
  • Examples:
    • Assigning Claude with Fable harness for iOS app UI development, Codex for image generation, and pi for complex planning.
  • Terminology:
    • crew-dispatch.json
    • harness
    • effort
    • grok-4.5
    • opus
    • fable
  • Why it matters: Optimizes token usage, speed, and output quality by matching tasks with the best-suited LLM and harness.

Quota Management with Quota-Axi

  • Explanation: Using quota-axi to monitor API subscriptions across Claude, Codex, Cursor, and Grok, automatically routing tasks to models with available quota.
  • Key claims:
    • Firstmate calls quota-axi to select the model with the most quota available, preventing hitting rate limits unexpectedly.
    • Automating quota checks saves time compared to manual monitoring.
  • Examples:
    • Checking quota status panel for Claude, Codex, Cursor, and Grok subscriptions.
  • Terminology:
    • quota-axi
    • quota
    • subscription
  • Why it matters: Ensures uninterrupted agent development by intelligently distributing workloads across available API quotas.

Backpass and Agent Memory

  • Explanation: Backpass analyzes past agent sessions using gradient descent principles to improve agent.md and cloud.md instructions for future runs.
  • Key claims:
    • Backpass closes the loop by having a human review past agent failures to improve agent instructions.
    • Automated collection of agent session transcripts across remote machines.
  • Examples:
    • Running backpass to collect transcripts across remote machines and dispatching a crewmate to work on it.
  • Terminology:
    • backpass
    • agent memory
    • gradient descent
    • transcripts
  • Why it matters: Continuously refines agent behavior and memory over time, reducing recurring mistakes.

Lavish Editor and Review Artifacts

  • Explanation: Lavish is an interactive HTML editor used by crewmates to present reports, design mockups, and decision points for captain review.
  • Key claims:
    • Interactive HTML artifacts allow the captain (human) to make decisions quickly without reading raw terminal logs.
    • Feedback given in Lavish is routed back to the appropriate agent crewmate.
  • Examples:
    • Reviewing bird variants and snow-capped mountain prototypes in Lavish editor.
  • Terminology:
    • Lavish
    • artifact
    • captain's call
    • decision
  • Why it matters: Simplifies human review and decision-making for complex multi-agent tasks using rich visual interfaces.

Scaling Up with Secondmates

  • Explanation: Scaling up management by spinning up secondmates to own different domains or projects, offloading orchestration from firstmate.
  • Key claims:
    • When firstmate gets too busy orchestrating tasks, secondmates are created to own specific app domains (e.g., Ship app, Edies Wallet app).
    • Secondmates act as firstmates for their respective sub-projects.
  • Examples:
    • Spinning up secondmates on Mac mini for separate app codebases.
  • Terminology:
    • secondmates
    • scaling up
    • Mac mini
    • headless
  • Why it matters: Enables massive horizontal scaling of agentic workflows across dozens of projects without hitting cognitive or computational bottlenecks.

Key Points

Talk to one agent, ship with a crew

  • Explanation: The developer only interacts with firstmate, which delegates and orchestrates all other agent crewmates in the background.
  • Evidence: Firstmate CLI commands and fleet registry listing 36 active projects.
  • Practical implication: Drastically reduces human cognitive load and context switching.

Calm Mode for Signal vs Noise

  • Explanation: Using the /calm command to hide verbose tool calls and terminal noise, showing only a floating boat indicator when agents are working.
  • Evidence: Toggling /calm mode in the terminal CLI.
  • Practical implication: Allows developers to focus on higher-level thinking and planning rather than watching terminal scrolling.

Policy-Driven PR Management (Yolo vs No-Mistakes)

  • Explanation: Each project has defined policies in projects.md specifying whether PRs require strict adversarial review (no-mistakes) or can be auto-merged (yolo).
  • Evidence: projects.md configuration showing policies like direct-PR+yolo or no-mistakes+yolo.
  • Practical implication: Automates routine merges for low-risk toys while applying rigorous validation for high-stakes code.

Human Attention is the Ultimate Bottleneck

  • Explanation: Trying to micro-manage every optimization leads to burnout. Focus human attention strictly on high-value architectural decisions.
  • Evidence: Discussion on context window limits (e.g., 500k auto-compact threshold) and avoiding manual micro-optimization.
  • Practical implication: Accept imperfect micro-optimizations and let automated pipelines handle routine tasks.

Frameworks, Models & Processes

Firstmate Agentic Workflow

  • How it works: Human talks to firstmate -> Firstmate reads intent and crew-dispatch rules -> Dispatches crewmates with specific harnesses and models -> Crewmates execute tasks, generate Lavish artifacts -> Human reviews artifacts and makes decisions -> Crewmates merge PRs.
  • Components:
    • Firstmate (primary orchestrator)
    • Crewmates (implementer agents)
    • Herdr (multiplexer)
    • Lavish (interactive UI review)
    • Backpass (memory improvement)
  • When to use: When managing multiple software projects and tasks simultaneously with AI coding agents.

Examples & Case Studies

Kun requested snow-capped mountain and bird variant prototypes for 'Fly With Me' game using Lavish editor.

  • Illustrates: Multi-agent prototyping and visual artifact review.
  • Lesson: Visual HTML artifacts make reviewing generative AI output extremely fast and intuitive.

Treehouse repo hook warning bug was reported on Discord, caught by bot, relayed to firstmate, investigated, and fixed via PR without human touching the code.

  • Illustrates: End-to-end asynchronous bug fixing across Discord, GitHub, and local agent session.
  • Lesson: Connecting chat platforms to agent orchestrators enables fully automated issue triage and resolution.

Actionable Takeaways

  • Immediate:
    • Set up a multiplexer like herdr to manage multiple agent sessions.
    • Use voice input for prompting agents to increase coding throughput.
    • Configure dispatch rules to route tasks to the most appropriate AI model.
  • Strategic:
    • Treat human attention and time as the ultimate bottleneck in agentic engineering.
    • Scale horizontally by delegating sub-projects to secondmate agents.
    • Build feedback loops (like backpass) to continuously improve agent memory.
  • Questions to investigate:
    • How to set up custom harness adapters for different LLM providers?
    • What is the optimal context window threshold for auto-compaction across different models?
    • How to configure robust adversarial validation pipelines for high-stakes repositories?

Claims Worth Verifying

  • Firstmate orchestrates multiple agent crewmates across 36+ projects simultaneously. (technical capability)
  • Backpass analyzes past agent session transcripts to improve agent.md memory via gradient descent principles. (methodology)

Notable Quotes

"I only talk to one agent, which is the firstmate. And the firstmate will orchestrate all the other agents for me." (at 1:45) "Our most precious resource right now, the biggest bottleneck across everything we do it with agents, is our attention and time." (at 56:40)

Compressed Summary

  • Use herdr tmux multiplexer for agent session management.
  • Firstmate orchestrates multiple agent crewmates via voice prompts.
  • Match task complexity and type to specific models using crew-dispatch.json.
  • Use Lavish for rich HTML artifact review and captain decisions.
  • Scale up by deploying secondmate agents to own separate sub-project domains.
  • Keywords: agentic, orchestration, multiplexer, multitasking, workflow
  • Core insight: High-throughput agentic engineering requires moving from manual micro-management to hierarchical agent orchestration where a primary agent manages specialized crewmates and interactive visual artifacts.

Core insights

7
Architecturemedium noveltymoderate evidence

Task-to-model routing is externalized into a declarative config (crew-dispatch.json) that maps task types to harness, model, and effort level, so the orchestrator selects execution resources without the developer choosing models per task.

Why it matters

Moves model/harness selection out of the developer's head and into a runtime policy layer, making cost/quality tradeoffs configurable and auditable rather than ad hoc.

Generalization

Any multi-model agent system can treat (task type -> model/harness/effort) as a routing table decoupled from task logic, enabling swapping models without touching orchestration code.

crew-dispatch.json configuration rules, which map specific task types, harnesses, models, and effort levels to different crewmates.
Open source video
Automated model selection based on task type prevents wasting expensive tokens.
Open source video
Mechanismhigh noveltymoderate evidence

Available subscription quota is treated as a first-class scheduling constraint: the orchestrator queries quota-axi and routes each task to the model with the most remaining quota.

Why it matters

Converts heterogeneous API subscription limits from an operational failure source into an input to task placement, avoiding mid-task rate-limit stalls.

Generalization

Agent schedulers can incorporate provider quota/capacity as a live resource dimension alongside capability and cost, similar to capacity-aware job scheduling in distributed systems.

Firstmate calls quota-axi to select the model with the most quota available, preventing hitting rate limits unexpectedly.
Open source video
Checking quota status panel for Claude, Codex, Cursor, and Grok subscriptions.
Open source video
Architecturemedium noveltymoderate evidence

Scaling is achieved by recursively replicating the orchestrator abstraction: when the top-level agent is saturated, 'secondmates' are spawned to own domains and themselves act as firstmates for their sub-projects.

Why it matters

Suggests horizontal scaling of agentic work is done by cloning the orchestration role per domain rather than strengthening a single centralized orchestrator.

Generalization

Hierarchical, self-similar orchestrators (manager-of-managers) may generalize to any workload where a single coordinator becomes a coordination bottleneck.

When firstmate gets too busy orchestrating tasks, secondmates are created to own specific app domains (e.g., Ship app, Edies Wallet app).
Open source video
Secondmates act as firstmates for their respective sub-projects.
Open source video
Mechanismhigh noveltyweak evidence

Agent instruction files (agent.md/cloud.md) are treated as learnable parameters: backpass mines past session transcripts across machines and revises those instructions, framed explicitly as gradient-descent-style optimization.

Why it matters

Positions prompts/config-as-code as an artifact subject to iterative optimization from observed failures, closing a feedback loop between runtime behavior and future instructions.

Generalization

Any agent harness with persistent instruction files can run an offline optimization pass over interaction logs to update its own configuration, independent of model weights.

Backpass analyzes past agent sessions using gradient descent principles to improve agent.md and cloud.md instructions for future runs.
Open source video
Running backpass to collect transcripts across remote machines and dispatching a crewmate to work on it.
Open source video
Practicemedium noveltymoderate evidence

Human-facing review is mediated by interactive HTML artifacts (Lavish) that agents generate to surface decisions, with feedback routed back to the originating crewmate rather than through raw terminal logs.

Why it matters

Reframes the human-agent interface as a structured decision protocol, reducing the review cost that caps how many parallel agents a human can supervise.

Generalization

Multi-agent systems that must keep a human in the loop benefit from typed, presentable decision artifacts instead of log streams as the review surface.

Interactive HTML artifacts allow the captain (human) to make decisions quickly without reading raw terminal logs.
Open source video
Feedback given in Lavish is routed back to the appropriate agent crewmate.
Open source video
Tradeoffmedium noveltymoderate evidence

Autonomy level for merges is set per project as policy (auto-merge 'yolo' vs adversarial-review 'no-mistakes'), allowing risk-tiered automation across a portfolio of repos rather than a single global setting.

Why it matters

Enables high throughput on low-risk work while reserving expensive verification for high-stakes code, which is otherwise a blanket policy decision.

Generalization

Autonomy and verification budgets should be declared per-workload/per-repo, not per-agent, mirroring blast-radius-based policy in deployment systems.

projects.md configuration showing policies like direct-PR+yolo or no-mistakes+yolo.
Open source video
Automates routine merges for low-risk toys while applying rigorous validation for high-stakes code.
Open source video
Mental Modelmedium noveltymoderate evidence

The system treats human attention, not compute, as the binding constraint: calm mode suppresses verbose tool output and micro-optimization is explicitly abandoned to preserve attention for architectural decisions.

Why it matters

Suggests throughput at scale is limited by human review capacity, so system design should optimize for minimizing human-visible signal rather than maximizing agent output.

Generalization

Agent platforms should be designed around an explicit human-attention budget, with noise suppression and triage as core features rather than UX polish.

Trying to micro-manage every optimization leads to burnout. Focus human attention strictly on high-value architectural decisions.
Open source video
/calm command to hide verbose tool calls and terminal noise, showing only a floating boat indicator when agents are working.
Open source video

Deep dives

5

Capability-constrained quota-aware scheduling for multi-model agent orchestrators

Research question

Does routing on provider quota as a first-class scheduler input reduce rate-limit-induced task failures relative to capability-only or round-robin routing, without degrading task quality beyond a defined threshold?

Why

Converts heterogeneous subscription limits from an operational failure source into a task-placement input; but greedily selecting the highest-quota model can override capability fit and trade quality for availability.

Firstmate calls quota-axi to select the model with the most quota available, preventing hitting rate limits unexpectedly.
Open source video
Checking quota status panel for Claude, Codex, Cursor, and Grok subscriptions.
Open source video
Source video

Instruction-file optimization from interaction transcripts (backpass)

Research question

Can treating agent.md/cloud.md as learnable parameters and applying an offline optimizer over past transcripts reduce recurring failure rates, and is the improvement distinguishable from human curation alone?

Why

The summary frames backpass as gradient-descent-style optimization while also involving human review; resolving this determines whether prompt/config-as-code becomes a measurable learnable artifact or remains a metaphor.

Backpass analyzes past agent sessions using gradient descent principles to improve agent.md and cloud.md instructions for future runs.
Open source video
Running backpass to collect transcripts across remote machines and dispatching a crewmate to work on it.
Open source video
Source video

Human-attention budget as the binding constraint in multi-agent throughput

Research question

Does explicitly measuring and budgeting human review attention improve end-to-end throughput more than adding additional agents or compute?

Why

If throughput is capped by human review capacity, system design should minimize human-visible signal and structure decisions rather than maximize agent output.

Trying to micro-manage every optimization leads to burnout. Focus human attention strictly on high-value architectural decisions.
Open source video
/calm command to hide verbose tool calls and terminal noise, showing only a floating boat indicator when agents are working.
Open source video
Source video

Cost/benefit of recursively cloning the orchestrator role (firstmate to secondmate)

Research question

At what domain count and concurrency level does adding a secondmate tier reduce end-to-end latency and error rate versus increasing single-orchestrator capacity, and what shared state/protocol is required?

Why

Recursive delegation may scale domains but introduces coordination and observability gaps between tiers; the tipping point is unmeasured in the summary.

When firstmate gets too busy orchestrating tasks, secondmates are created to own specific app domains (e.g., Ship app, Edies Wallet app).
Open source video
Secondmates act as firstmates for their respective sub-projects.
Open source video
Source video

Static versus learned task-to-model routing tables

Research question

Does periodically re-estimating crew-dispatch.json from outcome and cost logs outperform a static hand-written routing table on the cost-quality Pareto frontier?

Why

Routing config is a high-leverage artifact; if learned, errors can silently degrade cost or quality across all projects.

crew-dispatch.json configuration rules, which map specific task types, harnesses, models, and effort levels to different crewmates.
Open source video
Automated model selection based on task type prevents wasting expensive tokens.
Open source video
Source video

Article ideas

4

Stop Treating Model Quotas as an Ops Problem

Agent orchestrators should treat provider quota as a first-class scheduling constraint and solve capability-constrained quota-aware placement; naive highest-quota routing trades quality for availability, so the right design is constrained optimization, not greedy quota maximization.

Angle

Scheduling and optimization for multi-provider agent systems

Source video

Agent Instructions Are Parameters, Not Prose

If agent.md/cloud.md can be revised from interaction transcripts, they belong in an explicit optimization loop with measured failure deltas; without measurement, 'gradient descent' is only a metaphor.

Angle

Treating prompt/config-as-code as learnable artifacts

Source video

The Bottleneck Is You: Designing Agent Systems Around Human Attention

Multi-agent throughput scales only until human review saturates, so platforms should expose an attention budget, suppress verbose signals, and standardize decision artifacts as core features rather than UX polish.

Angle

Human-in-the-loop capacity planning

Source video

Manager-of-Managers: When Recursive Orchestration Beats a Bigger Orchestrator

Recursive firstmate/secondmate topologies pay off only with explicit shared state, reconciliation protocols, and cross-tier observability; without them, coordination overhead cancels the scaling benefit.

Angle

Hierarchical multi-agent architecture

Source video

Project ideas

4

Quota-Aware Capability-Constrained Router

new

A task router that optimizes placement under capability constraints and live quota signals reduces rate-limit-induced stalls by at least 30% versus capability-only routing, without more than a 5% relative quality regression on task-type benchmarks.

Proof of concept

Wrap a set of coding tasks with simulated provider quota APIs; route each task via capability-only, round-robin, highest-quota, and capability-constrained quota-aware policies.

Measurement

Stall/rate-limit rate, task success quality, token cost, end-to-end latency across routing policies.

Source video

Backpass Loop Harness

beyond-evals

One offline optimization pass that edits agent instruction files from historical transcripts reduces recurring failure rate on a held-out task set, with diminishing returns and measurable overfitting risk on repeated passes.

Proof of concept

Collect session transcripts, define a failure taxonomy, generate instruction-file edits, and rerun held-out tasks before/after each iteration.

Measurement

Held-out failure rate per iteration, instruction diff size, train-vs-held-out gap, human-review acceptance rate.

Source video

Attention-Budget Review A/B

movement-lab

Artifact-based review reduces human decision latency and error rate versus raw terminal logs, and latency gains diminish as artifact complexity increases.

Proof of concept

Present the same set of agent decisions to reviewers as structured HTML artifacts versus raw logs, randomizing order.

Measurement

Decision latency, decision error rate, self-reported cognitive load, review abandonment.

Source video

Risk-Tiered Merge Autonomy Simulator

gatehouse

Per-project autonomy policies (direct-PR+yolo vs no-mistakes+yolo) increase merged throughput per escaped defect compared with uniform auto-merge or uniform review, but only when risk classification accuracy exceeds a threshold.

Proof of concept

Simulate a portfolio of repos with injected defect rates and compare uniform and per-project merge/review policies.

Measurement

Merged changes per unit time, escaped defect rate, review cost, accuracy threshold where benefit reverses.

Source video

Architectural implications

4

Orchestration is centralized into a single conversational agent rather than distributed across developer-managed sessions.

Before

Developer juggles and switches between many terminal sessions and repos to direct agents.

After

Developer talks only to firstmate, which reads intent and dispatches crewmates across repos.

Consequence

A new orchestration layer becomes the critical path; its dispatch policy, quota awareness, and failure handling determine total system reliability.

Source video

Model selection becomes a runtime routing decision driven by task type and quota rather than a developer choice.

Before

Human picks a model per task, often defaulting to one expensive model.

After

crew-dispatch.json plus quota-axi select harness, model, and effort automatically.

Consequence

Routing config becomes a high-leverage artifact; errors in routing silently degrade cost or quality across all projects.

Source video

Human review shifts from reading raw execution logs to consuming generated artifacts with routed feedback.

Before

Human inspects terminal output to judge agent progress and correctness.

After

Agents emit Lavish HTML artifacts presenting decisions; human responds and feedback routes to the responsible crewmate.

Consequence

Agents must be able to produce structured decision surfaces, adding a presentation/summarization responsibility to the harness.

Source video

Coordination scales recursively by promoting a domain to its own orchestrator (secondmate).

Before

One orchestrator attempts to manage all projects.

After

Secondmates own domains and act as firstmates for their sub-projects, often on separate headless machines.

Consequence

Requires shared state/protocol between orchestrator tiers; otherwise recursive delegation introduces coordination and observability gaps.

Source video

Tradeoffs and failure modes

5

Per-project merge autonomy (yolo vs no-mistakes)

Benefit

Auto-merging low-risk changes raises throughput and avoids spending review budget on trivial work.

Cost or risk

Auto-merge policy can let defects into a repo; mis-classifying a high-stakes project as low-risk bypasses adversarial review entirely.

projects.md configuration showing policies like direct-PR+yolo or no-mistakes+yolo.
Open source video
Source video

Calm mode signal suppression

Benefit

Hides verbose tool calls and terminal noise so the human can focus on planning.

Cost or risk

Suppressing tool-call output reduces observability; silent agent misbehavior becomes harder to detect.

/calm command to hide verbose tool calls and terminal noise, showing only a floating boat indicator when agents are working.
Open source video
Source video

Autonomous memory/instruction optimization via backpass

Benefit

Continuously refines agent instructions from real transcripts, reducing recurring failures.

Cost or risk

Depends on human review of past failures and cross-machine transcript collection; instruction drift or overfitting to past sessions is a risk.

Backpass closes the loop by having a human review past agent failures to improve agent instructions.
Open source video
Source video

Ignoring micro-optimizations

Benefit

Preserves scarce human attention for architectural decisions and avoids burnout.

Cost or risk

Accepting imperfect micro-optimizations leaves systematic inefficiency (cost/latency) unaddressed indefinitely.

Trying to micro-manage every optimization leads to burnout. Focus human attention strictly on high-value architectural decisions.
Open source video
Source video

Quota-based routing

Benefit

Avoids hitting rate limits by sending work to models with the most remaining quota.

Cost or risk

Routing to the highest-quota model can override capability fit, potentially trading quality for availability; quota signal accuracy is assumed.

Firstmate calls quota-axi to select the model with the most quota available, preventing hitting rate limits unexpectedly.
Open source video
Source video

Open questions

5

Is dispatch policy in crew-dispatch.json statically authored, or is it updated from observed outcomes (as backpass implies for instructions)?

Why unresolved

The summary describes the config's contents but not whether routing rules are learned, tuned, or manually maintained.

Research direction

Prototype a routing table that is periodically re-estimated from task outcome/cost logs, and compare against a static hand-written table.

Source video

What concretely does 'gradient descent principles' mean for backpass — an actual optimizer over instruction text, or a human-curated review loop?

Why unresolved

The summary uses the gradient-descent framing while also stating the loop involves a human reviewing failures, leaving the mechanism ambiguous.

Research direction

Build a small harness that treats instruction files as parameters, runs a search/refinement procedure over transcripts, and measures failure-rate delta per iteration.

Source video

At what point does the firstmate/secondmate hierarchy stop helping and start adding coordination overhead?

Why unresolved

The summary asserts scaling benefit but reports no measurement of coordination cost or failure rates across tiers.

Research direction

Instrument a recursive orchestrator setup and measure end-to-end latency, error rate, and human interventions per added tier.

Source video

How reliable is the quota signal used for routing, and how is per-provider quota drained predicted?

Why unresolved

The summary states quota-axi selects the model with the most quota but gives no mechanism for measurement or prediction.

Research direction

Characterize quota API lag/error and test whether quota-based routing actually reduces rate-limit-induced task failures versus round-robin or capability-only routing.

Source video

Do interactive artifacts (Lavish) genuinely reduce human review time versus terminal logs, and at what artifact complexity do they become their own reading burden?

Why unresolved

The efficiency gain is asserted ('make decisions quickly') but not quantified.

Research direction

A/B a set of review tasks across artifact-based and log-based review, measuring decision latency and error rate.

Source video

Key claims

8
factualVerification needed

Firstmate is a single contact agent that orchestrates other agents (crewmates) across multiple repos.

Evidence

Firstmate acts as the primary contact agent that orchestrates all other agent crewmates.

Question

Does a single orchestrator agent reliably maintain correct state across many concurrent repos and sessions?

Source video
comparativeVerification needed

Different models excel at different task types (Fable for UI, Codex for image generation, Kimi/Astra for planning, Opus/Grok for bug fixes).

Evidence

Different models excel at different tasks (e.g., Fable for UI, Codex for image generation, Kimi/Astra for complex planning, Opus/Grok for bug fixes).

Question

Can these per-task model advantages be reproduced under controlled, same-task benchmarks?

Source video
causalVerification needed

Automatic task-type-based model selection prevents wasting expensive tokens.

Evidence

Automated model selection based on task type prevents wasting expensive tokens.

Question

How much token cost is actually saved, and does routing to cheaper models degrade task quality?

Source video
causalVerification needed

Firstmate queries quota-axi and routes tasks to the model with the most available quota, preventing unexpected rate limits.

Evidence

Firstmate calls quota-axi to select the model with the most quota available, preventing hitting rate limits unexpectedly.

Question

Does quota-aware routing measurably reduce rate-limit-induced stalls compared to capability-only routing?

Source video
causalVerification needed

Backpass improves agent.md and cloud.md instructions for future runs by analyzing past sessions using gradient descent principles.

Evidence

Backpass analyzes past agent sessions using gradient descent principles to improve agent.md and cloud.md instructions for future runs.

Question

Do instruction updates derived from past transcripts reduce subsequent failure rates, and is the process actually an optimizer?

Source video
factualVerification needed

Secondmates act as firstmates for their sub-projects, enabling scaling across separate app codebases.

Evidence

Secondmates act as firstmates for their respective sub-projects.

Question

What protocol/state sharing is required between orchestrator tiers, and where does it break down?

Source video
factualVerification needed

The speaker manages 36+ projects simultaneously without cognitive overload.

Evidence

he manages 36+ projects simultaneously without cognitive overload

Question

Is 'without cognitive overload' substantiated by any measurement, or is it a subjective self-report?

Source video
causalVerification needed

Interactive HTML artifacts let the human make decisions quickly without reading raw terminal logs.

Evidence

Interactive HTML artifacts allow the captain (human) to make decisions quickly without reading raw terminal logs.

Question

Does artifact-based review measurably reduce decision latency/error versus reading logs?

Source video

Connections

5