article
Treat Continuity as a First-Class Product Capability
Durable events, state, memory, goals, attention, and decisions let products preserve coherent progress without turning context into unchecked authority.
3m read
Building thoughtful software with AI.Notes from the systems behind the work.
Durable events, state, memory, goals, attention, and decisions let products preserve coherent progress without turning context into unchecked authority.
Let models propose task meaning while deterministic contracts decide which capabilities may run.
Represent each request as a typed semantic task envelope containing objective, continuity, constraints, entities, evidence needs, capability inputs, dependencies, and completion criteria; let a model propose the plan, then validate it against executable contracts and fall back safely when planning fails.
Turn reusable guidance into a learning product with time-aware attribution, evidence-linked evaluation, versioned amendments, and preserved lineage.
Evidence-backed triggers, bounded procedures, portable sources, scenario tests, progressive discovery, and active curation turn workflow lessons into durable shared capability.
Privacy-aware traces, structured facets, complementary evaluation, human review, and measured improvements turn completed AI-assisted work into shared capability.
Scoped contracts, least-privilege evidence, deterministic validation, human review, sandbox evaluation, and maintained releases turn repository judgment into reusable agent capability.
Semantic checkpoints, distinct progress dimensions, evidence-led adaptation, and direct outcome verification make long-work evaluation useful and humane.
Capability-based readers, semantic normalization, stable identity, governed provenance, and conformance tests turn heterogeneous agent histories into one extensible activity model.
A Dec 11 implementation makes an AI review operation a first-class run: a trigger writes a typed database record and returns a run ID quickly, background execution receives a sandbox and durable log, and a status reader exposes the run state, recommendation, scores, findings, timestamps, and duration.
A small supervisor-led routing layer lets an AI service accept one request, hand it to the right specialist, and return a structured result.
A shared metadata contract lets an AI workbench discover tool modules, explain them in the settings surface, and route enabled actions without maintaining a second hand-written index.
A growing local toolset becomes safer to extend when discovery, argument transport, path authority, output limits, and failure semantics are shared by one explicit invocation contract.
A security finding should not jump straight to a change. A tenant-scoped, rate-limited, observable pipeline can generate a proposed remediation while keeping every mutation reviewable.
Long-running agent requests can return immediately without becoming opaque: persist the analysis and run first, then stream bounded evidence through a tenant-scoped background lifecycle.
An agent can analyze the wrong codebase or lose its history when repository selection lives only in a request. Durable analysis records bind organization, repository, actor, status, evidence, and agent execution so self-improvement remains reviewable and tenant-safe.
An agent that analyzes a codebase and proposes new capabilities can become an unsafe mutation loop without explicit state, dry-run defaults, credential scoping, durable evidence, and human review. A control plane turns self-improvement into a bounded workflow.
A repository-aware improvement loop analyzes a codebase, compares its existing agent capabilities with its actual stack, and turns the gaps into tailored skills and reviewable implementation changes.
A model-backed desktop chat gained an incremental reply path that updates the assistant bubble as text arrives, then settles it into a completed message with useful metadata.
A reliable multi-agent workflow is a state machine, not a string of prompts. This case study shows how typed state, routing, bounded iteration, and synthesis make agent orchestration testable.
A robust AI controller can use rejection feedback to improve an answer while preserving the user's original authority, completion criteria, and evidence obligations.
A canonical guide to evaluating agentic work from privacy-preserving evidence, ordered gates, and explicit limits rather than raw activity transcripts.
A canonical guide to describing, discovering, selecting, and safely invoking tools in a local coding agent.
A source-neutral account of how typed event registries and constrained hook runners make local agent lifecycle behavior testable without confusing observation with permission.