What survives interruption, restart, protocol change, or a client disappearing?
Find durable session identity, serialized state, replay rules, expiry, cancellation, reconciliation, and the point where an unknown effect becomes safe to retry.
A project appears here because the boundary is relevant to its documented or reported architecture. The evidence label says how far this Atlas has checked the mechanism.
The runtime can inspect pending input, resolve permissions and tool mode, run hooks, and add higher-authority context before a model invocation is assembled.
BeforeAgent and related hook boundaries let the runtime attach delimited context or block a run before the next agent step.
Release and issue trails distinguish account, network, proxy, authentication, and CLI behavior when the implementation surface is not equally open.
Provider-specific session identifiers and expiry behavior are kept in typed session state instead of being flattened into one generic transport assumption.
Recovery rebuilds the client registry as a coherent unit instead of leaving a partially alive tool menu in place.
Control-flow interrupts are separated from ordinary failures so a retry layer does not consume the operator's stop or handoff signal.
Thread-local approval state narrows the risk that consent for one task becomes ambient authority for another.
Diff evidence becomes trustworthy only when the runtime records enough repository state to distinguish agent output from a guessed comparison.
An adapter can turn a public API description into model-visible tools while keeping schema validation, credentials, execution, and result handling separate.
A sign-in flow remains pending with a claim link and explicit state instead of turning an operator action into an arbitrary elapsed-time failure.
ACP support is tied to release evidence so compatibility is more than a capability claim in documentation.
Compatibility code and tests keep older serialized message history usable while the wire and extension model changes.
A retained run keeps changes beside the working tree until a reviewer selects, applies, releases, or discards the output once.
The project repeatedly merges Codex upstream while owning release, installer, cache, documentation, and harness changes for a different product bet.
The project documents an automation mode in which permission prompts are disabled, making the trust placed in the surrounding CI or sandbox environment explicit.
The documented daemon mode serves one agent session to multiple clients over HTTP and server-sent events, turning client identity and session ownership into protocol questions.
The project documents human-in-the-loop suspension backed by storage so a workflow can pause and later resume from durable execution state.
The project documents a Plan mode for investigation and strategy, an Act mode for execution, per-action approvals, optional auto-approval, diffs, and reversible checkpoints.
The project documents a local Rust agent that connects provider choices and external capabilities through standard extension contracts instead of baking every integration into one model path.
The framework presents agent dependencies, tool inputs, and structured results through Python types and validation boundaries.
The repository separates agent construction from explicit multi-agent workflow and deployment APIs in two language ecosystems.
The same control surface can connect to an agent running directly on a host, in Docker, on a VM, or in a cloud service; the README explicitly warns that an unsandboxed server has full filesystem access.
The agent can complete a web action while still violating the user's intended destination, account, or consequence; evaluation must inspect the effect, not only task completion.
Crew and task abstractions make responsibilities explicit, but dependable delegation also requires lifecycle, authority, result, cancellation, and evidence semantics.
A direct message resolves to a personal scope, a group message to a group scope, and a channel to a channel scope; the resolver then mounts organization state read-only and the active scope read-write before a harness runs.
The launch command configures distinct protocol, context, compaction, state, and permission adapters for Claude Code, Codex, OpenCode, OpenClaw, Hermes, and Pi instead of treating one compatible endpoint as complete agent equivalence.
A BEHAVIOR.md names recurring expected conduct and its failure boundary; prompts, skills, traces, rubrics, and scorers can change around it without silently becoming the source of truth.
The monolith writes a future wake epoch, the always-alive dispatcher turns it into a trajectory step, and a backoff policy decides how soon the agent should think again based on whether the last wake produced visible work, only a thought, an idle, or an error.
The bundled code, dependency, and cloud-posture workflows name concrete scanners, require coverage disclosures when a tool is missing, and reserve fixes or cloud changes for reviewable code paths.
Compatibility fixtures keep persisted histories usable across a protocol rewrite. alejandro-ao/tau
Codex Keeps the Prompt After You Interrupt ItPending input remains attributable after an interruption instead of disappearing into runtime state. openai/codex
OpenWork Gives MCP Failures a Useful AddressA failed session becomes recoverable when the runtime can identify the exact connection that died. different-ai/openwork
September 5, 2026 — Crush’s Hyper Prism work carries the routed model from response headers to a durable turn record—the provenance pattern builders need whenever a router may change what actually served the request. charmbracelet/crush
API Discovery: An MCP Hint Can Feed a Human Approval MapSeptember 4, 2026 — LangChain's beta MCPAdapter carries a server's tool metadata into an agent runtime, giving builders a practical way to pause risky calls for a human—without mistaking a hint for authorization. langchain-ai/langchain
Hermes Decides That a Preference Can Have a HomeSeptember 4, 2026 — Hermes is redrawing a deceptively important line: a correction learned while doing one task belongs with that task's skill, while persistent memory is for facts that should matter everywhere. NousResearch/hermes-agent
Before Codex Trusts a Workspace, It Stops Letting the Workspace Choose Its HelpersSeptember 3, 2026 — A merged Codex change moves automatic pre-trust helper lookups to known system locations, sanitizes plugin-sync Git, and makes diagnostics inspect a workspace-influenced path without running what they find. openai/codex, google-gemini/gemini-cli
API Discovery: Let an Agent Notice When Its Reasoning Was DroppedSeptember 3, 2026 — Pydantic AI now carries an Anthropic thinking-block fallback through later turns—but the useful operator move is to expose the loss, then fix the unstable prompt prefix that caused it. pydantic/pydantic-ai
Builder Work: An Approval Must Point to This Exact InvocationSeptember 3, 2026 — Microsoft Agent Framework’s Copilot-assisted approval repair shows the craft worth stealing: authorize one stable local occurrence, not a provider’s reusable tool-call label. microsoft/agent-framework
A Child Agent’s Approval Request Shouldn’t Hide in Another TabSeptember 2, 2026 — Openwork now brings a child session’s blocked permission into the parent task; fresh Goose state-machine work shows why seeing the ask is only half the control. different-ai/openwork, aaif-goose/goose
Builder Work: A Socket Path Is Not a Permission BoundarySeptember 2, 2026 — Qwen Code’s Claude-assisted inbox hardening shows the reusable craft: treat an endpoint and its credential as one capability, then erase both before a child process can inherit them. QwenLM/qwen-code
This is a mechanism comparison, not a feature score. “Reporting underway” means the Atlas has not verified the relevant path; it does not mean the project lacks it.