Agent Loop
Own the repeated transition from user intent to model response, tool effects, updated context, and an explicit stop reason.
- Lessons
- 3
- Snapshots
- 6
- Claims
- 6
- Experiments
- 0
Mechanisms connect lessons, agent snapshots, claims, and experiments. Similar interfaces are not treated as identical designs without evidence.
Own the repeated transition from user intent to model response, tool effects, updated context, and an explicit stop reason.
Create a version-bound baseline before edits, review a scoped state delta, and restore only the explicitly selected paths when recovery is needed.
Expose the harness loop, sessions, events, approvals, and configuration to terminal, IDE, desktop, or automation clients through a stable transport boundary.
Allocate finite input space across instructions, conversation, tools, memory, and output reserve.
Replace an over-budget model-visible history with a smaller projection while preserving enough state, provenance, and recent work to continue safely.
Translate provider deltas and harness transitions into ordered, replayable events.
Decide which tool actions may run, which require a human or policy decision, and which must be constrained by an operating-system boundary.
Load and execute hooks, plugins, skills, custom tools, or provider adapters without collapsing extension code into the core loop.
Discover, order, bound, and attribute project instructions before model execution.
Store selected, source-attributed facts outside the immediate context and retrieve only relevant records for the current turn.
Constrain outbound destinations, protocols, resolution, redirects, and credential flow independently from whether a tool is semantically approved.
Enforce filesystem, process, and related capability limits below semantic tool policy so compromised or incorrect code cannot exceed its execution envelope.
Reject absolute paths and traversal before a file tool touches the filesystem.
Bind explicit trust to a stable workspace identity and decide which project inputs may become instructions instead of ordinary untrusted data.
Classify failures, bound retries, avoid duplicate side effects, and preserve enough state for a safe continuation or explicit pause.
Rebuild derived session state from an append-only event log and create lineage-preserving branches without rewriting the parent run.
Delegate bounded work into isolated contexts, control capabilities and lifecycle, and merge results without confusing ownership or workspace state.
Validate, authorize, execute, normalize, and record model-requested tool calls.
Bound oversized tool results while preserving provenance, size, and a useful preview.
Expose file, command, and edit effects through explicit, inspectable tool boundaries.