Reference architecture patterns for production AI agent systems. Not better models. Better engineering discipline: context architecture, execution guardrails, and memory infrastructure that actually works.
Document input flows through the Aho-Corasick engine to a knowledge graph, filtered by role, and dispatched to hook, REST, or WASM runtime targets.
The harness, not the model, determines agent success. LangChain's coding agent jumped from 52.8% to 66.5% on Terminal Bench 2.0 by changing zero model parameters.
Deciding what information the model sees at each step. Context is a scarce resource: a 1M token window with 800K of noise performs worse than 200K with 150K curated signal. Structured compaction before the window fills.
Enforcing what the model can and cannot do. Risk-tiered CI with machine-readable contracts: high-risk paths get policy gates, browser evidence, and security scans. Low-risk paths get fast-tracked.
Ensuring the model learns from its own history. Failed commands captured automatically, corrections applied, patterns graduated to rules. Cross-session search prevents repeating solved problems.
Bigger context windows do not solve the problem. They make it worse without proper architecture.
By 100K tokens, the window is 60% noise: old file reads, superseded search results, abandoned reasoning paths. Adding more capacity without compaction makes performance worse, not better.
System prompt says "use pnpm." CLAUDE.md says "run tests before committing." README says "use npm." By 200K tokens, instructions compete with 150K of conversation history and lose.
Long-running agents accumulate so many intermediate steps that they lose sight of the original objective. The model optimises for local coherence rather than global progress toward the actual goal.
Framework sprawl versus disciplined harness engineering.
| Capability | Reference Architecture | Framework + Cloud | Raw LLM API |
|---|---|---|---|
| Context compaction | |||
| Risk-tiered CI gates | |||
| Cross-session learning | |||
| Role-based knowledge graphs | |||
| Multi-agent coordination | |||
| Strategic drift detection | |||
| Data stays local | |||
| Vendor lock-in |
Full support Partial Not available
85% planning, 15% code. Moving the hardest thinking into representations that still fit inside model context windows is the entire game.
Explain goals to frontier models. Establish problem boundaries and success criteria before any design work begins.
Request independent designs from multiple models. Diverse reasoning paths surface blind spots that a single model misses.
Create a "best-of-all-worlds" hybrid plan. Merge the strongest elements via structured diff format.
4-5+ rounds in fresh conversations. Plans reach 3,000-6,000+ lines. Each round sharpens architecture and edge cases.
Plan Space: 1x Rework CostTransform markdown into executable work units with full dependency graphs and embedded context. Every piece of reasoning travels with the task.
Task Space: 5x Rework Cost4-6+ refinement cycles. "Check tasks N times, implement once." Transition when whole-workflow questions stabilise.
Deploy 5-15 coordinated agents. Stagger starts by 30+ seconds to prevent thundering herd. All agents are generalists, instantly replaceable.
Code Space: 25x Rework CostHuman manages flow, unblocks stuck tasks. Cross-agent review every 30-60 minutes alternating random exploration and systematic integration checks.
Self-review, cross-review, testing, polish. Strategic drift detection: "Do we actually have the thing we are trying to build?"
From the core thesis to production patterns.
Articles on harness engineering, context management, and production agent architecture.
The knowledge graph engine behind the reference architecture. Local-first, sub-millisecond, Rust/WASM.
Technical reference for the 52 Rust crates, deployment guides, and integration patterns.
Open source under Apache 2.0. Contributions welcome. All code reviewed by the reference architecture patterns.
52 crates, 5-10ns search latency, 15MB footprint. The engine underneath the architecture.
Production SaaS with connectors, LLM proxy, and agentic skills for enterprise deployment.
Type at least 2 characters.