Loops that call tools, the protocol that exposes them, and what to keep in context.
Sign in to track your progress across sections.
43 articles
Listed inWhat are AI Agents?Agents and toolson
Anthropic on the difference between workflows and agents, five composable patterns, and why most teams should start with the simplest thing that works.
Listed inMulti-Agent SystemsAgents and toolson
Lilian Weng's survey of planning, memory, and tool use — still the clearest map of the agent design space.
Listed inGoogle ADKAgents and toolson
Google's ADK: agent definitions, tools, sessions, and the built-in evaluation harness.
Listed inClaude Agent SDKAgents and toolson
Writing the agent loop is straightforward; the harness supplies what surrounds it. Context management, sessions, permissions, and the file and shell tools an agent needs are what adopting the SDK gives you.
Listed inManual ImplementationAgents and toolson
The agent loop in forty lines, why a failing tool returns its error as a result, and what a framework is doing on top.
Listed inTools & Function CallingAgents and toolson
The model chooses a tool from its description alone, and every declaration is billed on every request whether or not it is called. Assume any tool with a side effect will be invoked twice.
Listed inModusAgents and toolson
The repository is archived and read-only, last pushed in September 2025. The idea worth keeping is the sandboxed WebAssembly module, which bounds what an agent can touch in a way importing a library does not.
Listed inGoogle ADKAgents and toolson
`SequentialAgent`, `ParallelAgent` and `LoopAgent` make the workflow patterns deterministic classes rather than behaviour you hope for, which suits a pipeline with one judgement call in it. Model-agnostic across five language bindings, with evaluation in the box.
Listed inWhat are AI Agents?Agents and toolson
Anthropic's line is who decides what happens next: your code, or the model. The five workflow patterns that sit between one call and an agent, and the three costs the loop brings with it.
Listed inVertex AI Agent BuilderAgents and toolson
Google's managed half, where grounding, scaling and availability are configured rather than assembled. IAM, VPC Service Controls and data residency are usually what decides it, and what you give up is visibility into the retrieval pipeline.
Listed inOpenAI AgentKit & Agents SDKAgents and toolson
Three approaches form a ladder rather than alternatives: a hosted Agents API, an SDK that runs inside your application, and the raw Responses API. Guardrails are a validation step you configure, not an enforcement boundary the model cannot cross.
Listed inMulti-Agent SystemsAgents and toolson
Four topologies, and costs that do not appear in a token count: latency is the slowest path, and four agents at 90% each is nearer a 65% system. The documented failures were delegation quality rather than architecture, which means the fix is the brief.
Listed inAgent Use CasesAgents and toolson
An agent costs roughly four times a chat interaction, so the test is whether the path is unknown in advance and whether success is checkable. Software tasks pass both; anything with a fixed procedure passes neither.
Listed inData LayerAgents and toolson
The normative source: lifecycle, capability negotiation, transports, and the exact shape of tools, resources, and prompts.
Listed inModel Context Protocol (MCP)Agents and toolson
Hosts, clients and servers, and the three primitives a server exposes: tools, resources and prompts. Includes a working TypeScript server in about forty lines, and the mistakes that make a server unusable in practice.
Listed inModel Context Protocol (MCP)Agents and toolson
The announcement that started the ecosystem — why a shared protocol beats N×M bespoke integrations.
Listed inTransport LayerAgents and toolson
stdio or streamable HTTP, at revision 2026-07-28, which removed sessions, GET streams and resumability. What moving off stdio forces you to add.
Listed inMCP ServerAgents and toolson
Tools are model-controlled, resources are application-controlled. That column decides which primitive a capability belongs to.
Listed inMCP HostAgents and toolson
The six responsibilities the specification gives the host, why one client per server is the isolation boundary, and what happens when a server asks you for a completion.
Listed inConnect to a Local ServerAgents and toolson
Connecting over stdio is a process launch, not a connection. Absolute paths, the rule against printing to stdout, and where the logs are.
Listed inBuilding an MCP ServerAgents and toolson
The input schema tells the model what to send as well as validating what arrives. Returning tools in a stable order protects callers' prompt caches, and an error that names the rule lets the model correct itself.
Listed inData LayerAgents and toolson
The current MCP revision is stateless: server/discover replaces the handshake, and every request carries its own protocol version and capabilities in _meta. Notifications are opt-in now, and sampling and logging are both deprecated.
Listed inMCP ClientAgents and toolson
The per-server connection the host spawns to speak the protocol.
Listed inConnect to a Remote ServerAgents and toolson
Transport, authentication and trust all change at once. A remote server can alter its tool descriptions without a deploy on your side, which means it can change your prompt, and a shared service credential leaves it unable to tell your users apart.
Listed inModel Context Protocol (MCP)Agents and toolson
A standard wire format so tools work with any model host.
Listed inBuilding an MCP ClientAgents and toolson
The protocol is handled by every SDK, so the real work is translation: MCP tool definitions into your provider's format, namespacing two servers that both offer `search`, and deciding which tools the model sees at all.
Listed inLong Context ProcessingAgents and toolson
The paper behind the U-shaped recall curve: models attend to the start and end of a long context and lose the middle.
Listed inWhat is a Context Layer?Agents and toolson
One place in the codebase that decides what goes into every prompt. The limit that binds is the model's attention rather than the size of the context window, so the layer curates what it sends instead of filling the space available.
Listed inContext EvaluationAgents and toolson
Wrong answer or missing context? Precision has a reference-free variant you can run today; recall always needs ground truth.
Listed inContext EngineeringAgents and toolson
Five things make up the payload, and each is a decision with its own failure. Tool definitions are billed on every call whether used or not.
Listed inLong Context ProcessingAgents and toolson
The curve is U-shaped and the middle is the weak spot, even for models trained for long context. Where to put what matters.
Listed inAtlanAgents and toolson
The context an assistant needs for warehouse data is metadata, not rows. Certification is what makes it trustworthy enough to act on.
Listed inContext IsolationAgents and toolson
Running a sub-task in its own context window keeps its exploration out of the main conversation, which can return a short result from a long search. The same separation is a normative rule in the Model Context Protocol, and it costs the caller information the sub-task never reports.
Listed inState & Historical ContextAgents and toolson
Twenty turns into a booking, six settled facts are what the model needs and the transcript is an inefficient way to carry them. Trimming a turn per request can also cost more than sending everything, because it breaks the cache.
Listed inContext SourcesAgents and toolson
A request that looks like one prompt is usually four or five fetches, each with its own latency, freshness and trust level. Giving every source a token budget at assembly makes the total knowable before the request goes out.
Listed inDataHubAgents and toolson
Once a retrieval corpus is real, three questions arrive that nothing in an AI stack answers: what is in it, where each piece came from, and whether you may use it. Lineage answers which parts of your index a broken upstream source affected.
Listed inMemory SystemsAgents and toolson
A memory directory an agent fills without discrimination is worse than none, because a wrong note gets cited on every later task. Facts that persist and progress on unfinished work age differently, so they belong in separate files.
Listed inPrompt vs Context EngineeringAgents and toolson
Wording the ask vs assembling everything the model needs to answer it.
Listed inContext SecurityAgents and toolson
Prompt injection is text arriving that should not steer the model; this is data sitting in the window that should not be there at all. The tenant filter belongs inside the retrieval query, and the payload you assembled now exists in four other systems.
Listed inContext Failure ModesAgents and toolson
Poisoning, distraction, confusion and clash are four measured ways a window degrades, and shortening the context helps only two of them. One study found a 39% drop from delivering the same information across several turns instead of at once.
Listed inMulti-Agent Context SharingAgents and toolson
Separate context windows are the reason to run several agents, since each one condenses what it read before reporting back. Multi-agent systems use about 15 times the tokens of a chat, which is the figure any architecture has to justify.
Listed inContext CompactionAgents and toolson
Compaction replaces older turns with a summary Claude writes on the server, and the swap that follows has two mistakes that raise no error. Images and documents inside the summarized range do not survive it.
Listed inExternal MemoryAgents and toolson
Claude's memory tool runs client-side, so your application executes every file operation and /memories is a prefix you map onto real storage. That makes path traversal your bug to prevent, and the documented example handlers skip the validation on purpose.