Orchestration & Control
42 篇文章
Action-Selector Pattern
Untrusted input can hijack an agent's reasoning once tool feedback re-enters the context window, leading to arbitrary, harmful actions.
emergingAgent-Driven Research
Traditional research methods often lack the ability to adapt search strategies based on emerging results, limiting efficiency and potential discoveries.
establishedAgent Modes by Model Personality
Different AI models have fundamentally different personalities and working styles. Treating all models the same—expecting them to work identically—leads to suboptimal outcomes. Users expect a consiste…
emergingAutonomous Workflow Agent Architecture
Complex, long-running engineering workflows traditionally require extensive human oversight and intervention. Tasks like model training pipelines, infrastructure configuration, and multi-step deployme…
establishedBurn the Boats
In fast-moving AI development, holding onto features or workflows that are "working fine" prevents teams from fully embracing new paradigms. The comfort of existing functionality becomes an anchor tha…
emergingContinuous Autonomous Task Loop Pattern
Traditional development workflows require constant human intervention for task management: - **Manual Task Selection**: Developers spend time deciding what to work on next from todo lists - **Context…
establishedCustom Sandboxed Background Agent
Off-the-shelf coding agents (e.g., Devin, Claude Code, Cursor) are either: - **Too generic** - Not deeply integrated with company-specific dev environments, tools, and workflows - **Vendor-locked** -…
emergingDiscrete Phase Separation
When AI agents attempt to simultaneously research, plan, and implement solutions, context contamination occurs. Competing priorities within a single conversation degrade output quality as the agent st…
emergingDisposable Scaffolding Over Durable Features
In a field where foundation models improve dramatically every few months, investing significant engineering effort into building complex, durable features *around* the model is extremely risky. A feat…
best-practiceDistributed Execution with Cloud Workers
Single-session AI agent execution cannot scale to meet enterprise team demands. Complex projects require multiple simultaneous code changes across different parts of the codebase, but coordinating mul…
emergingDual LLM Pattern
A privileged agent that both sees untrusted text **and** wields tools can be coerced into dangerous calls.
emergingExplicit Posterior-Sampling Planner
Agents that rely on ad-hoc heuristics explore poorly, wasting tokens and API calls on dead ends.
emergingFactory over Assistant
The "assistant" model—working one-on-one with an agent in a sidebar, watching it work, ping-ponging back and forth—limits productivity and scalability. As models become more autonomous and capable, th…
emergingFeature List as Immutable Contract
Long-running agents exhibit several failure modes when tasked with building complete applications: - **Premature victory declaration**: Agent declares "done" after implementing a fraction of requirem…
emergingHybrid LLM/Code Workflow Coordinator
LLM-driven workflows are **non-deterministic**—even well-crafted prompts can produce unpredictable results. For some tasks (e.g., adding emoji to Slack messages based on PR status), occasional errors …
proposedInference-Time Scaling
Traditional language models are limited by their training-time capabilities. Once trained, their performance is essentially fixed, regardless of how much compute is available at inference time. This m…
emergingInitializer-Maintainer Dual Agent Architecture
Long-running agent projects face distinct failure modes at different lifecycle stages: - **Project initialization** requires comprehensive setup: environment configuration, feature specification, tes…
emergingInversion of Control
Traditional "prompt-as-puppeteer" workflows force humans to spell out every step, limiting scale and creativity.
validated-in-productionIterative Multi-Agent Brainstorming
For complex problems or creative ideation, a single AI agent instance might get stuck in a local optimum or fail to explore a diverse range of solutions. Generating a breadth of ideas can be challengi…
experimental-but-awesomeLane-Based Execution Queueing
Traditional agent systems serialize all operations through a single execution queue, creating bottlenecks that limit throughput. Concurrent execution is desirable but risky: - **Interleaving hazards*…
validated-in-productionLanguage Agent Tree Search (LATS)
Current language agents often struggle with complex reasoning tasks that require exploration of multiple solution paths. Simple linear approaches like ReACT or basic reflection patterns can get stuck …
emergingLLM Map-Reduce Pattern
Injecting a single poisoned document can manipulate global reasoning if all data is processed in one context.
emergingMulti-Model Orchestration for Complex Edits
A single large language model, even if powerful, may not be optimally suited for all sub-tasks involved in a complex operation like multi-file code editing. Tasks such as understanding broad context, …
validated-in-productionOpponent Processor / Multi-Agent Debate Pattern
Single-agent decision making can suffer from: - **Confirmation bias**: Agent finds evidence supporting initial hypothesis - **Limited perspectives**: One context window misses alternative approaches …
emergingOracle and Worker Multi-Model Approach
Relying on a single AI model creates a trade-off between capability and cost. High-performance models are expensive for routine tasks, while cost-effective models may lack the reasoning power for comp…
emergingParallel Tool Call Learning
Agents often execute tool calls sequentially even when they could run in parallel: - **Unnecessary latency**: Sequential calls add up when tool execution time dominates inference time - **Inefficient…
emergingConditional Parallel Tool Execution
When an AI agent decides to use multiple tools in a single reasoning step, executing them strictly sequentially can lead to significant delays, especially if many tools are read-only and could be run …
validated-in-productionPlan-Then-Execute Pattern
If tool outputs can alter the *choice* of later actions, injected instructions may redirect the agent toward malicious steps.
emergingPlanner-Worker Separation for Long-Running Agents
Running multiple AI agents in parallel for complex, multi-week projects creates significant coordination challenges: - **Flat structures** lead to conflicts, duplicated work, and agents stepping on e…
emergingProgressive Autonomy with Model Evolution
Agent scaffolding built for older models becomes unnecessary overhead as models improve: - **Prompt bloat**: System prompts accumulate instructions that newer models don't need - **Over-engineered fl…
best-practiceProgressive Complexity Escalation
Organizations deploy AI agents with overly ambitious capabilities from day one, leading to: - Unreliable outputs when agents tackle tasks beyond current model capabilities - Failed implementations th…
emergingRecursive Best-of-N Delegation
Recursive delegation (parent agent -> sub-agents -> sub-sub-agents) is great for decomposing big tasks, but it has a failure mode: - A single weak sub-agent result can poison the parent's next steps …
emergingSelf-Rewriting Meta-Prompt Loop
Static system prompts become stale or overly brittle as an agent encounters new tasks and edge-cases. Manually editing them is slow and error-prone.
emergingSpecification-Driven Agent Development
Hand-crafted prompts or loose user stories leave room for ambiguity; agents can wander, over-interpret, or produce code that conflicts with stakeholder intent.
proposedStop Hook Auto-Continue Pattern
Agents complete their turn and return control to the user even when the task isn't truly done. Common scenarios: - Code compiles but tests fail - Changes made but quality checks haven't passed - Feat…
emergingSub-Agent Spawning
Large multi-file tasks blow out the main agent's context window and reasoning budget. You need a way to delegate work to specialized agents with isolated contexts and tools.
validated-in-productionSubject Hygiene for Task Delegation
When delegating work to subagents via the Task tool, empty or generic task subjects make conversations: - **Untraceable**: Cannot identify what a subagent was working on - **Unreferencable**: Cannot …
emergingSwarm Migration Pattern
Large-scale code migrations are time-consuming when done sequentially: - **Framework upgrades** (e.g., testing library A → testing library B) - **Lint rule rollouts** across hundreds of files - **API…
validated-in-productionThree-Stage Perception Architecture
Complex AI agents often struggle with unstructured inputs and need a systematic way to process information before taking action. Without a clear separation of concerns, agents can become monolithic an…
proposedTool Capability Compartmentalization
Model Context Protocol (MCP) encourages "mix-and-match" tools—often combining private-data readers, web fetchers, and writers in a single callable unit. This amplifies the lethality of prompt-injectio…
emergingTool Selection Guide
AI agents often struggle to select the optimal tool for a given task, leading to inefficient workflows. Common anti-patterns include: - Using `Write` when `Edit` would be more appropriate - Launching…
emergingTree-of-Thought Reasoning
Linear chain-of-thought reasoning can get stuck on complex problems, missing alternative approaches or failing to backtrack.
established