AI Agent、MCP、工具调用、多 Agent 协作
This collection features articles and community discussions in the Agent 工程 space, focusing on real-world engineering scenarios, tool selection, practical paths, and risk boundaries.
The page is continuously updated and serves as a learning roadmap, technology selection reference, and technical review entry point.
`resources/updated` only signals that a URI may have changed; it carries no content version. This article uses discovery generations, subscription leases, and read-digest reconciliation to handle lost notifications, out-of-order updates, and resource-set changes during pagination.
`modelPreferences` does not specify a model, and Sampling results should not reach the Server without review. This article freezes the routing policy, actual model, tool loop, and two-stage request/result consent into a single delivery receipt.
When an MCP Server connects to production systems, the hardest problem is not obtaining a token but preventing identity and permission boundaries from being conflated. This article presents a practical approach to layered authentication and authorization, token exchange, auditing, and failure handling.
An SSE disconnect, HTTP 404, and DELETE do not signify the same kind of termination. This article binds the MCP session, protocol version, original stream, and Last-Event-ID to prevent cross-stream replay, session hijacking, and duplicate side effects.
MCP tools must handle model retries, network instability, and long-running downstream jobs. This article presents a complete design for timeout budgets, retry conditions, idempotency keys, cancellation semantics, and observability metrics to prevent duplicate writes and indefinite waits.
An MCP tool's input and output schemas form a long-lived contract among the model, client, and business service. This article covers parameter design, error models, pagination, compatibility, and testability so a growing tool catalog remains usable.
Routing tasks to the “strongest Agent” by role name ignores tool permissions, version compatibility, and real-time capacity. This article uses versioned capability descriptors, tiered admission control, and deadline-aware load shedding to protect interactive and critical workloads.
Drawing a DAG does not make it recoverable. This article separates dependency satisfaction, node attempts, artifact commits, and graph revisions, using event replay and a derived ready set to prevent missed runs, premature runs, or duplicate downstream release after a restart.
After multiple agents return results, concatenation or voting loses evidence and applicability conditions. This article designs candidate-result lineage, deterministic conflict classification, arbitration records, and atomic publication so merged conclusions can be recomputed and revoked.
When multiple agents concurrently modify shared plans, preferences, or conclusions, simple last-write-wins handling loses facts. This article uses versioned commands, per-key ordering, and field-level merge policies to control concurrency.
When multiple Agents compete for the same work, a TTL-based claim helps with reclamation but cannot stop an old executor from committing late. This article uses monotonic fencing, attempt identity, and side-effect records to establish provable task ownership.
When Agents wait on one another for results, locks, approvals, or capacity, per-task timeouts merely prolong deadlocks. This article builds a versioned wait graph, detects strongly connected cycles, and releases resources through verifiable cancellation convergence.