AI agents: orchestration platforms
10
Glossary and sources
5 min read · Chapter 10 of 10
Words that mean something else here
| Term | What you'd assume | What it means here |
|---|---|---|
| Agent | Software that acts on its own | Anything from a chatbot to a workflow; engineers mean the model picks the next step |
| Autonomous | No human involved | Usually fewer than 10 steps, then a person |
| Memory | The agent remembers you | The context, a session store, extracted facts or training; only facts persist per user |
| Context window | What the model uses | The model's maximum; products cap it lower, and models use less well |
| Tool call | The model did something | The model asked; your runtime acted and owns the side effect |
| Guardrail | A security control | Often a bypassable classifier; only a deterministic policy check is a control |
| Sandbox | Isolated | A shared-kernel container, gVisor, a microVM, or just "no network" |
| MCP server | A remote service | Often a local process with your full privileges |
| Registry | A vetted catalogue | Verified names and metadata; nobody scans the code |
| Zero data retention | Nothing stored | Per endpoint; stateful features and some top models are excluded |
| Deprecated | Gone | Usually still working, with a date set; in Azure's API, already retired |
| Deployer, provider | DevOps words | EU legal roles; an enterprise can become the provider |
| Indemnity | Cover for what the agent does | Copyright claims on outputs only |
What surprised me
The metered call. In telco I could multiply minutes by the price on a rate deck. Here every lap re-sends the whole conversation, so the meter speeds up the longer the call lasts, and an uncapped loop costs fifteen times a normal run.
Remittance data. In order to cash, I learned to match data that arrives separately from the money. Here the data can give orders: an email the agent reads is an instruction it may follow, and no parameterized query stops it.
Approval as the control. In procure to pay, an approver checks a purchase order against an invoice. Here the approver sees a prompt with little context and, in one lab's data, clicks yes about 93% of the time, and an approval nobody reads ends up as little more than a log entry.
The merchant-locked card. In spend management, a card locked to one merchant capped the damage at one merchant and one amount. A scoped token caps which host the agent can reach, but a fooled agent can still do anything the token allows there.
Someone else's calendar. In cross-border payouts, holidays and cut-offs belonged to the destination. Here the whole product lives on someone else's calendar: the model I tested against retires on 45 days' to six months' notice, and its replacement counts tokens differently.
Sources
Read on October 2, 2026 unless dated; "search result" means seen only as a search snippet.
Protocols, standards and frameworks
- MCP: versioning, 2026-07-28 changelog, release post, authorization, security practices, tools, registry; A2A and its 2026 blog archive; Agentic AI Foundation and the Linux Foundation's announcement
- RFC 8693; IETF AIMS, transaction tokens and cross-app access; OpenID Foundation on AuthZEN; OpenTelemetry GenAI; OWASP LLM Top 10, 2026 edition and prompt injection; MITRE ATLAS; NIST AI RMF
- Agent payments: FIDO Alliance on AP2; Stripe on ACP
Governments, regulators and courts
- EU AI Act Art. 25 and Art. 26; GPAI Code signatories (Jul 2026); CNIL on erasure
- UK AISI (2026); NCSC; CISA (May 2026); NIST agent initiative, NCCoE, CAISI (Mar 2026)
Labs
- Anthropic: pricing, deprecations, retention, data residency, rate limits, tool use, MCP connector, usage policy, prompt injection, constitutional classifiers, engineering posts on building effective agents, its multi-agent research system, advanced tool use and approval data
- OpenAI: pricing, deprecations, data controls, agent guide, Agents SDK, function calling, tool search, shell tool, MCP, usage policies (search result), usage API (search result)
- Google: pricing, models, terms, usage policy, Interactions API, function calling, managed agent with MCP, its agent kit's sessions, a forum thread on an embedding model's retirement; Google DeepMind on prompt injection defenses (v1)
- Meta: agent security; Mistral: models; xAI: agent terms (search result)
- Size and moves (main players): Bloomberg Government on Anthropic's run-rate (Aug 2026); Axios on OpenAI's run-rate (Sep 2026) and how the two labs count revenue (search results); Alphabet's Q2 2026 results; Reuters on Meta's paid API (Jul 2026) and SpaceX buying xAI (Feb 2026); Siècle Digital on Mistral's revenue target (Jan 2026, search result)
- Self-serve tiers, startup programs and enterprise deals (read October 11, 2026 unless dated): OpenAI rate limits; Anthropic's startup program (Oct 2026); OpenAI's startup credits via partners (search result); OpenAI Frontier via AI Business (Feb 2026, search result)
Clouds
- AWS: AgentCore pricing, harness, limits, security, model lifecycle, data protection
- Microsoft: responsibility split, model retirements and the retirement schedule, agent design patterns, Foundry data privacy, Entra agent identities (search result)
- Google Cloud: agent identity, design patterns, data governance, Agent Builder release notes (search result)
- Results (main players): Amazon (Jul 2026); Microsoft's FY26 Q4 call (Jul 2026, transcript); Google Cloud is in Alphabet's results under Labs
- Startup credits and commitments: AWS Activate credits for Bedrock models (May 2026); Google for Startups AI tier and a Microsoft for Startups guide (search results); Claude in Microsoft Foundry (search result); Microsoft 365 E7 via Office Watch (search result)
Research
- Measuring Agents in Production, turn control, MAST, scaling agent systems, τ-bench, METR, benchmark checklist, AgentPoison, adaptive attacks
- Harnesses position paper, context length, RouteLLM, guardrail bypass, injection competition
Analysts, surveys, filings, incidents
- KPMG (Sep 2026), McKinsey (Nov 2025), Gartner (Jun 2025, also on "agent washing"), a framework vendor's survey (Jun 2026, vendor)
- Budgets and prices by size: KPMG's Q3 2026 AI Pulse report; a16z CIO survey (Jan 2026; a16z invests in OpenAI); Ramp's AI Index via a news write-up (Sep 2026); Gartner on customer service agents (Mar 2025, search result); Epoch AI on price trends (search result)
- Salesforce and ServiceNow results (company-reported); Salesforce and Intercom pricing
- Main players, company results and releases (2026 unless dated): Databricks, Snowflake, Glean, Elastic, MongoDB, Datadog, ClickHouse on Langfuse, Okta, Cloudflare, Workday, SAP via Futurum; Palo Alto Networks on Protect AI and its 10-K (2025); CrowdStrike's 10-K on Pangea (all company-reported; most seen as search results or reposts)
- OECD incident records on Replit and Antigravity; AI Incident Database on AWS and Cursor (search results)
- NVD: CVE-2025-3248, CVE-2025-32711, CVE-2025-49596, CVE-2025-6514, CVE-2025-54135, CVE-2025-53773, CVE-2026-30623
Law firms and press (search results unless noted)
- Sidley, McDermott, Orrick, Debevoise, McCarthy Tétrault (all fetched)
- Duane Morris, Cuatrecasas, Latham, DLA Piper, Norton Rose Fulbright, WSGR, hh-law, Lawfare, FPF
- GeekWire and Retail Insight Network on Amazon v. Perplexity; TechCrunch on Meta and the Code, Anthropic hiring an eval team, insurers and OpenAI on prompt injection; AI Business and VKTR on OpenAI's acquisitions; The Register; BNN Bloomberg on Deloitte; Malay Mail on fake court citations; Big I and beinsure on insurance
- Compute and data access (2026): The Decoder and Fast Company Middle East on rationing (Apr 2026, fetched); an advisory blog on OpenAI's capacity reservations; Hunton on Slack's API terms
- Funding (main players): SiliconANGLE on LangChain and Braintrust, The Next Web on Temporal, Bloomberg Law on Decagon and Let's Data Science on Sierra (all fetched); Sifted on n8n, HPCwire on Arize, Unite.AI on E2B
Vendor and secondary sources (small companies kept generic)
- A graph framework's interrupt and persistence docs; a workflow-engine vendor's idempotency note
- LangSmith pricing (Oct 2026)
- A security firm's tool-poisoning proof of concept; a security vendor's write-up of the fake npm server; Cloud Security Alliance notes on MCP marketplaces and EU standards; a security newsletter; a security news site on the archive wipe; an MCP CVE aggregator I couldn't verify
- An EU enforcement tracker; a careers blog on the Workday case; a Copilot licensing guide and a licensing site (search results)
Field Guides are learning notes, not legal or compliance advice. Rules and fees change; check the cited primary sources before you act on anything here.