FIIT.AI BlogMultilingual SEO notes distilled from public X discussions about Codex, Agent Harness, memory, tools, safety, and production agent engineering.
Codex's docs dedicate a section to 'Agent configuration', centered on AGENTS.md, Subagents, Speed, and Rules. AGENTS.md is a project-level instruction file placed in the repo root that the agent reads before touching code — so build commands, test conventions, and naming rules are stably inherited instead of re-explained in every chat.
Codex's Permissions section lists five layers — Profiles, Sandboxing, Auto-review, Agent approvals & security, and Internet access. It's not an all-or-nothing switch but a layered mechanism that tightens agent permissions: sandboxes isolate the execution environment, auto-review gates code changes before they land, and internet access is controlled separately.
Codex's Environments section lists Modes, Local environments, Cloud environment, and Git worktrees. The same agent can run in your local environment, be handed to a cloud environment, or use Git worktrees so each task proceeds in its own branch worktree — 'where the work happens' is itself a configurable engineering decision.
a2aproject/A2A (26k stars, hosted by the Linux Foundation and contributed by Google) defines the Agent2Agent open protocol, addressing how agents from different companies, frameworks, and servers collaborate 'as agents' rather than being called 'as tools'. It is built on JSON-RPC 2.0 over HTTP(S), discovers capabilities via Agent Cards, supports synchronous request/response, SSE streaming, and async push, and ships six language SDKs.
The A2A protocol lists 'Preserve Opacity' among its design goals: agents can collaborate without sharing internal memory, proprietary logic, or specific tool implementations. Capabilities are exposed via the Agent Card as a 'minimal contract', while internal implementation stays a black box — solving both 'can cooperate' and 'dare to cooperate' for cross-company agents.
OpenAI lists Compaction among the API's 'core concepts', alongside Conversation state, Streaming, and Counting tokens. It answers the question every multi-turn agent eventually faces: as conversation history grows toward the context-window limit, how do you compress it without losing critical information, so long tasks keep running.
The OpenTelemetry specification (a CNCF project, open-telemetry/opentelemetry-specification) defines traces, metrics, and logs as one cross-language, cross-implementation standard, versioned under Semantic Versioning 2.0. Codex's docs also list Tracing and Observability and usage as standalone modules — agent observability is moving from 'dig through logs after the fact' to 'structured, correlatable tracing'.
OpenAI's Agents SDK lists Guardrails as a standalone module alongside Orchestration and Results and state; the docs also include Safety best practices, Red teaming, and Safety checks / Safety classifiers. Safety is no longer just a pre-launch spot check but a first-class citizen embedded in the agent's build and runtime.
Codex's What's new shows: on September 21–25, GPT-6 Sol and GPT-6 Luna began rolling out in Codex at lower token prices than GPT-5.6; GPT-5.5 is scheduled to retire from ChatGPT, ChatGPT Work, and Codex on October 14, 2026. Model generational turnover is now a regular rhythm in Codex, which means a concrete upgrade window for workflows that depend on it.
Codex's docs list Multi-agent as a core API concept and pair it with Orchestration and Sandbox agents in the Agents SDK. OpenAI's approach isn't 'one giant agent does everything' but splitting work across specialized agents, organized into pipelines or hierarchies through orchestration and handoffs.