Find a concept.
27 concepts
- Agent harness
The surrounding software that runs a model’s work loop, supplies tools, manages context and records progress. The harness and the model contribute different parts of the overall agent behaviour.
Explore this in a development workflow →- Agentic development
A software workflow in which AI agents can use tools and perform bounded work toward a goal, rather than only suggesting text.
Explore this in a development workflow →- Multi-agent orchestration
Coordination of multiple agents through task ownership, dependencies, state and handoffs. In a coding project, it connects planning, implementation and review.
Explore this in a development workflow →- Control plane
The coordination layer that decides what may run and records workflow state. Model execution performs the reasoning and implementation.
Explore this in a development workflow →- Task graph / DAG
A representation of tasks and dependencies. A directed acyclic graph describes a dependency structure without circular dependencies; not every project workflow is automatically a DAG.
Explore this in a development workflow →- Git worktrees
Multiple working trees attached to a Git repository. They provide separate checkouts, but do not by themselves reconcile conflicting product decisions.
Explore this in a development workflow →- Backpressure
Limiting or redirecting incoming work when a downstream stage cannot keep up. A growing review queue is a reason to inspect review capacity before adding more coders.
Explore this in a development workflow →- Work stealing
A scheduling pattern where an available worker takes suitable queued work from elsewhere. Task readiness, permissions and ownership still need to be valid.
Explore this in a development workflow →- Model routing
Selecting a model for a particular task using requirements, capabilities and allowed resources. A cheap token price alone does not establish the best choice.
Explore this in a development workflow →- Reasoning effort
A model-specific control over reasoning. It is not a universal measure across providers or a direct count of agents.
Explore this in a development workflow →- Capability floor
The minimum acceptable model or reasoning capability for a task. Sensitive or unknown work should not be downgraded solely to use a cheaper option.
Explore this in a development workflow →- Context engineering
Selecting, structuring and maintaining the information an AI needs: requirements, relevant files, instructions, decisions and evidence.
Explore this in a development workflow →- Context compaction
Reducing conversation history into a smaller representation. It can consume usage and lose detail, so durable requirements should remain available separately.
Explore this in a development workflow →- Prompt caching
Provider-specific reuse of previously processed input. Eligibility and billing depend on the provider; it is different from saving a local conversation.
Explore this in a development workflow →- Agent handoffs
Transfer of a bounded task and its relevant state to another responsible worker or stage. Useful handoffs identify the version, requirements and next action.
Explore this in a development workflow →- Agent skills
Reusable instructions or procedures with a defined scope. A skill should not silently expand the permission granted for a particular task.
Explore this in a development workflow →- Subagents and helpers
Delegated workers used for a bounded task. Provider-native subagents and workspace-managed helpers may expose different control, status and usage information.
Explore this in a development workflow →- Quality gates
Defined checks that a delivery must satisfy before advancing. Their value depends on the exact version, relevant tests and handling of unknown results.
Explore this in a development workflow →- Evals / evaluation harness
Repeatable tasks, starting conditions and acceptance checks used to compare a model or workflow. Representative outcomes matter more than an isolated successful demo.
Explore this in a development workflow →- Human-in-the-loop
A workflow that includes human judgment at meaningful decision points. It does not require a person to relay every routine agent message.
Explore this in a development workflow →- Spec-driven development
Development organized around explicit requirements and acceptance criteria. Natural-language conversation can still be the way the specification evolves.
Explore this in a development workflow →- BYOK and connected accounts
Bring your own key typically means using an API key from your provider. Connecting a signed-in account is a different access path and can have different billing.
Explore this in a development workflow →- MCP / Model Context Protocol
A protocol for connecting AI applications with tools and context. Supporting a protocol does not by itself guarantee that every tool or provider is supported by a product.
Explore this in a development workflow →- Idempotency
A property where repeating the same operation does not create an additional effect. It is useful for retries, but only for the operations designed to provide it.
Explore this in a development workflow →- Agent observability
Being able to inspect task state, ownership, attempts, outcomes and usage where measured. An animated indicator alone is not evidence of useful progress.
Explore this in a development workflow →- Checkpoint and resume
Recording enough durable state to continue after interruption. Resuming work still requires checking uncertain external effects and the current code.
Explore this in a development workflow →- Agent swarm
A broad term for multiple cooperating agents. The label says little about coordination quality unless responsibilities, limits and verification are explained.
Explore this in a development workflow →
Read the primary documentation.
For implementation details, start with the Git worktree manual, the Model Context Protocol introduction and Anthropic’s explanation of agent harnesses. Vraido’s guides explain how these broader concepts relate to coordinated software work.