Glossary
Key terms and definitions for agentic engineering
Core Concepts
Foundational patterns for agentic systems
-
Agentic swarm
A coordinated group of specialized agents that ships work with explicit handoffs, verification gates, and build receipts.
Also: swarm, agent-swarm
-
Gated multi-agent flow
A multi-stage agent workflow where each stage must pass verification gates before work can move forward.
Also: gated-flow, gated-workflow
-
Trust boundary
The line where you stop trusting model narration and require tool-verified evidence (tests, diffs, logs) before accepting progress.
Also: evidence-boundary, verification-boundary
Agent Roles
Specialized agent types and their responsibilities
-
Explore agent
A read-mostly agent role that maps a codebase or problem space, surfaces risks, and proposes a bounded plan without making changes.
Also: explorer-agent
-
Task agent
An implementation-focused agent role that takes a scoped task and ships a small, reviewable change with tests and receipts.
Also: implementation-agent
-
Wave of agents
A coordination pattern: dispatch many small agents in parallel (then in waves) so each owns one issue, with a coordinator integrating results.
Also: agent-wave
Guardrails
Safety mechanisms and validation patterns
-
Build receipt
A structured, machine-verifiable record of what an agent ran and what happened (commands, versions, exit codes, diffs, artifacts).
Also: receipt, ci-receipt
-
Context starvation
A failure mode where an agent lacks the constraints, history, or code references needed to make correct changes—leading to confident wrong work.
Also: context-limiting, goldfish-strategy
-
Mechanical counting
Using deterministic tools (scripts, calculators, hashes) for anything that must be exact instead of trusting model-generated counts.
Also: kernel-truth, tool-verification
-
Mutation on diff
Run mutation testing only against changed code to prove tests would fail if the new logic were wrong—high signal without full-suite cost.
Also: diff-mutation-testing
-
Oppositional validation
Deliberately adversarial validation: a separate agent or check tries to disprove a change, not approve it.
Also: adversarial-validation
-
Schema gravity
The tendency for systems to drift toward what their schemas and contracts validate—useful when schemas match reality, painful when they don't.
Also: schema-contract-gravity
-
Shadow fork
An isolated branch or environment where agents can run commands and try changes safely without contaminating main or production credentials.
Also: isolated-run, run-sandbox
Operations
Running agentic systems in production
Failure Modes
Common pitfalls and anti-patterns
-
Process confabulation
A failure mode where an agent narrates progress it didn't actually perform (tests run, files changed, PR opened).
Also: fake-progress, narrated-success
-
Reward hacking
Optimizing for a proxy metric or gate (tests, lint, score) in a way that makes the metric look good while degrading the real outcome.
Also: metric-gaming