Classification · 2026-10-03
Agent types
What kind of agent is doing the work, which existing Depot jobs it needs, and which failure modes to watch. Needs point at JTBD ids and free-resource kinds already in the catalog. This page does not change Depot Picks, price sort, or the free index.
GET /api/agent-classes · 15 classes · failures · case studies
Coding agent
coding
Edits a repository, runs commands in a workspace, and checks the result against tests or a diff.
Failure modes: tool-loop, hallucinated-tool-args, partial-tool-success, collateral-side-effects, weak-verification, eval-contamination, benchmark-protocol-mismatch, runaway-spend, silent-truncation, non-determinism, schema-drift, context-overflow, instruction-drift
- Patch a failing test and run that test in a sandbox
- Review a diff without rewriting unrelated files
- Stop when the same command repeats with the same output
Studies: SWE-bench: Can Language Models Resolve Real-World GitHub Issues?, Introducing SWE-bench Verified, Cognition SWE-bench technical report, Building effective agents, TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks, AgentBench: Evaluating LLMs as Agents, AppWorld: A Controllable World of Apps and People for Benchmarking Interactive Coding Agents
/agents/kit/coding · GET /api/kit?class=coding
Research agent
research
Gathers sources for an open-ended question and returns an answer that names those sources.
Failure modes: prompt-injection, stale-retrieval, tool-loop, runaway-spend, lost-in-the-middle, ungrounded-answer, instruction-drift, inter-agent-misalignment, context-overflow, policy-conflict, silent-truncation
- Answer a question with the URLs that support each claim
- Say when the fetched pages do not contain the answer
- Split a question across workers and check that each result matches its assignment
Studies: GAIA: a benchmark for General AI Assistants, Building effective agents, How we built our multi-agent research system, TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks, Lost in the Middle: How Language Models Use Long Contexts
/agents/kit/research · GET /api/kit?class=research
Browser agent
browser
Completes a task by reading and operating a website in a browser, then stops before a consequential submit unless confirmation exists.
Failure modes: prompt-injection, tool-loop, gui-grounding-miss, partial-tool-success, non-determinism, confirmation-skip, hallucinated-tool-args, collateral-side-effects, over-refusal, instruction-drift, benchmark-protocol-mismatch
- Find a stated price on a page and quote the URL
- Fill a form on a test site and stop before submit
- Refuse to continue when the page text tries to replace the user's task
Studies: WebArena: A Realistic Web Environment for Building Autonomous Agents, VisualWebArena, Computer-Using Agent, Introducing Operator, TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks, WorkArena: How Capable Are Web Agents at Solving Common Knowledge Work Tasks?, Prompt injection: What's the worst that can happen?
/agents/kit/browser · GET /api/kit?class=browser
Computer-use agent
computer-use
Operates a desktop or mobile GUI, not only a single site API, and checks that the control it used is the one it intended.
Failure modes: gui-grounding-miss, prompt-injection, collateral-side-effects, partial-tool-success, tool-loop, confirmation-skip, benchmark-protocol-mismatch
- Change one setting and verify the resulting file
- Stop if the click opened a different app than the plan named
Studies: VisualWebArena, OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments, OSWorld project site: verified upgrade, Computer-Using Agent, Developing a computer use model
/agents/kit/computer-use · GET /api/kit?class=computer-use
Voice agent
voice
Hears or speaks on a call or a realtime channel, and confirms account-changing actions before they run.
Failure modes: silent-truncation, identity-confusion, over-refusal, non-determinism, partial-tool-success, confirmation-skip, runaway-spend
- Transcribe a caller and repeat the requested change before acting
- Hand off when the transcript is cut off
Studies: τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains
/agents/kit/voice · GET /api/kit?class=voice
Support agent
support
Answers a customer from the operator's policy and escalates when sources disagree or the action is consequential.
Failure modes: policy-conflict, ungrounded-answer, stale-retrieval, identity-confusion, prompt-injection, over-refusal, confirmation-skip, non-determinism, weak-verification, schema-drift
- Answer a policy question by quoting the canonical page
- Show both pages when they disagree instead of picking one
- Do not send a refund without an approval id
Studies: τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains, WorkArena: How Capable Are Web Agents at Solving Common Knowledge Work Tasks?, Moffatt v. Air Canada, 2024 BCCRT 149, Why we rebuilt LangChain's chatbot
/agents/kit/support · GET /api/kit?class=support
Data and ETL agent
data-etl
Moves or reshapes records between files, tables, and documents, and reports rows that did not match the schema.
Failure modes: schema-drift, partial-tool-success, silent-truncation, stale-retrieval, hallucinated-tool-args, collateral-side-effects
- Load a table and list rows that failed the schema
- Parse a document and mark pages the OCR step did not return
Studies: dbt Cloud scheduled jobs failed on an unexpected Databricks adapter argument
/agents/kit/data-etl · GET /api/kit?class=data-etl
RAG agent
rag
Answers from a retrieved corpus and says when the corpus does not support the claim.
Needs
Jobs: rag-enterprise, rag-memory, hybrid-search-rag, rag-rerank, prototype-rag, knowledge-graph-rag, late-interaction-retrieval, embeddings, managed-vector, rag-evals
Free kinds: free_embedding, free_vector, free_rerank, free_data
Tags: rag, vector
- Passage ids on every factual sentence
Failure modes: stale-retrieval, lost-in-the-middle, silent-truncation, eval-contamination, ungrounded-answer, context-overflow, policy-conflict
- Answer only from retrieved passages
- Say the corpus has no support instead of filling the gap
Studies: Lost in the Middle: How Language Models Use Long Contexts, MemGPT: Towards LLMs as Operating Systems, Why we rebuilt LangChain's chatbot
/agents/kit/rag · GET /api/kit?class=rag
Multi-agent orchestrator
orchestrator
A lead assigns subtasks to workers and accepts a result only when it matches the assignment.
Failure modes: inter-agent-misalignment, tool-loop, runaway-spend, identity-confusion, partial-tool-success, weak-verification, instruction-drift, hallucinated-tool-args, observability-gap
- Delegate two searches and reject a worker result that answers a different question
- Stop the fan-out when the tool-call budget is spent
Studies: Building effective agents, How we built our multi-agent research system, TheAgentCompany: Benchmarking LLM Agents on Consequential Real World Tasks, Why Do Multi-Agent LLM Systems Fail?
/agents/kit/orchestrator · GET /api/kit?class=orchestrator
Eval and judge agent
eval-judge
Scores a trace or a patch against a rubric, without the gold label in its prompt, and records the harness version.
Failure modes: eval-contamination, non-determinism, weak-verification, observability-gap, benchmark-protocol-mismatch, inter-agent-misalignment
- Score a trace against a rubric that does not include the gold answer
- Refuse to rank two runs that used different harness versions
Studies: SWE-bench: Can Language Models Resolve Real-World GitHub Issues?, Introducing SWE-bench Verified, Cognition SWE-bench technical report, OSWorld project site: verified upgrade, AgentBench: Evaluating LLMs as Agents, Why Do Multi-Agent LLM Systems Fail?, LangSmith incident on May 1, 2025
/agents/kit/eval-judge · GET /api/kit?class=eval-judge
Scheduled and ops agent
scheduled-ops
Runs a job on a queue or a clock, is safe to retry, and alerts when the job or its traces go dark.
Failure modes: runaway-spend, partial-tool-success, tool-loop, observability-gap, non-determinism
- Retry a failed job without applying the side effect twice
- Alert when the night job's trace count falls
Studies: LangSmith incident on May 1, 2025
/agents/kit/scheduled-ops · GET /api/kit?class=scheduled-ops
Personal assistant
personal-assistant
Acts for one person across sessions, keeps that person's memory separate, and waits before sending or booking.
Failure modes: identity-confusion, context-overflow, confirmation-skip, stale-retrieval, prompt-injection, instruction-drift
- Draft a calendar hold and wait for confirmation
- Remember a preference for this user only
Studies: GAIA: a benchmark for General AI Assistants, Introducing Operator, AppWorld: A Controllable World of Apps and People for Benchmarking Interactive Coding Agents, MemGPT: Towards LLMs as Operating Systems, Prompt injection: What's the worst that can happen?
/agents/kit/personal-assistant · GET /api/kit?class=personal-assistant
Sales and outreach agent
sales-outreach
Drafts outreach from sourced fields about a person or account, and does not send until a person approves.
Failure modes: identity-confusion, confirmation-skip, runaway-spend, stale-retrieval, policy-conflict, ungrounded-answer, prompt-injection
- Draft an email from cited fields and leave it unsent
- Stop a sequence when the recipient has already replied
Studies: FTC press release: Verkada commercial email and the CAN-SPAM Act
/agents/kit/sales-outreach · GET /api/kit?class=sales-outreach
Security review agent
security-review
Reads a change or a trace for secrets, overly broad permissions, and untrusted text copied into a tool prompt. It does not produce an attack.
Failure modes: prompt-injection, over-refusal, eval-contamination, weak-verification
- List secrets and broad permissions in a diff
- Flag a prompt that concatenates fetched page text into a tool instruction, without quoting an attack
Studies: Developing a computer use model, Prompt injection: What's the worst that can happen?
/agents/kit/security-review · GET /api/kit?class=security-review
Local and offline agent
local-offline
Runs without calling a hosted model API, using local weights, a local index, or an air-gapped serve path.
Needs
Jobs: air-gapped-model-serve, open-model-inference, self-host-vector, self-host-browser, wasm-edge-small-models, on-device-edge-batch, edge-sqlite-agent-db
Free kinds: open_weights, local_runtime, free_vector, free_sandbox
Tags: oss, edge
- No network call on the answer path
- Say when the local corpus has no support
Failure modes: silent-truncation, context-overflow, schema-drift, non-determinism, stale-retrieval, lost-in-the-middle
- Answer from a local index and say when it has no passage
- Refuse to call a hosted model when the task is marked offline
Studies: MemGPT: Towards LLMs as Operating Systems
/agents/kit/local-offline · GET /api/kit?class=local-offline