--- name: harness-frontier-diagnosis-2026-06-02 description: Frontier digest anchored on FH's 3-layer identity (Control Tower · Frontier→Org Propagation · AI Collaboration Guide) + Core Axis. External AI/harness-engineering signal from 2026-06 translated into per-identity strengthening candidates, with simplicity guards. type: frontier-diagnosis date: 2026-06-02 engine: websearch tags: [frontier, identity, harness-engineering, multi-agent, a2a, mcp, context-engineering, observability, v2-paper] --- # Harness Frontier Diagnosis — 2026-06-02 > Identity ② asset (`harness_frontier_diagnosis_*.md`). Collects the global AI/harness-engineering > frontier and **translates it for FH operations** — anchored on FH's three identities + Core Axis. > Engine: WebSearch (no `ANTHROPIC_API_KEY`; outbound curl blocked → forced downgrade per `frontier-digest` skill). ## FH Identity Anchor (from CLAUDE.md §Identity) | # | Identity | One-line role | |---|---|---| | ① | **Control Tower** | Command HQ that coordinates all connected projects | | ② | **Frontier → Org Propagation** | Absorb global frontier thinking, translate it into org language | | ③ | **AI Collaboration Guide** | Accumulate/distribute token-efficiency + dialogue methodology | | Axis | **Harness Engineering (How)** | The 6-axis methodology that realizes the three above | --- ## Frontier Highlights (2026-06) **1. "Harness Engineering" named the 4th paradigm of AI engineering.** The arc prompt → context → harness is now an explicit industry framing: *"Agents aren't hard; the Harness is hard."* The widely-cited claim is that **~65% of enterprise AI failures trace to harness defects** — Context Drift, Schema Misalignment, State Degradation — not model capability. This is direct external validation of FH's whole thesis (`meta_harness_engineering_definition.md`, `fh_ecosystem_positioning.md`). → **Core Axis**. **2. Agent interoperability standardized: A2A "Agent Cards" + MCP registry under Linux Foundation.** A2A standardizes how agents *discover* each other's capabilities (Agent Cards); MCP launched a community server registry (Nov 2025). Production topology data: orchestrator-worker is ~70% of deployments, but **centralized multi-agent coordination carries ~+285% token overhead** and the practical team size is **3–4 agents** before coordination cost dominates. → **① Control Tower**. **3. Observability is the bottleneck for self-improving harnesses (Agentic Harness Engineering).** AHE's central claim: *agents cannot reliably improve a black-box harness* — the evolution loop needs the harness's components, experiences, and decisions to be observable and verifiable. Eval-driven: the 2026 Coding Agent Index benchmarks **model+harness pairs**, not models alone. Context research adds the "lost in the middle" effect (10–30% accuracy drop on mid-context information) and the hierarchical-context remedy (L1 always-on / L2 session / L3 on-demand) + prompt compression. → **② Frontier Propagation** + **③ AI Collaboration Guide**. --- ## Per-Identity Strengthening Candidates ### ① Control Tower | Candidate | Frontier basis | FH hook | |---|---|---| | **Machine-readable `agent-card`-style capability registry** for the FH agents (capability / input-output contract, synced to actual file counts) | A2A Agent Card = the discovery standard | Closes the "count drift / no canonical registry" gap already flagged in `fh_ecosystem_positioning.md` | | **Coordination-overhead budget** (then in `context-bridge-dispatch`, merged into `agent-composer` Step 3-a on 2026-06-02): parallel-fan-out cap (3–4) + capability-aware routing | Centralized = +285% tokens; team size caps at 3–4 | `plugins/fh-meta/skills/agent-composer` | ### ② Frontier → Org Propagation | Candidate | Frontier basis | FH hook | |---|---|---| | Add a **harness-defect taxonomy axis** (Context Drift / Schema Misalignment / State Degradation) to structural diagnosis | "65% of AI failures = harness defects" | `plugins/fh-meta/skills/harness-doctor` | | Add **observability / eval hooks** to the evolution loop so self-improvement is glass-box, not black-box (eval-driven, model+harness benchmarking style) | AHE: observability is the self-improvement bottleneck | `plugins/fh-meta/skills/harvest-loop`, `harness-doctor` | ### ③ AI Collaboration Guide | Candidate | Frontier basis | FH hook | |---|---|---| | Formalize **L1/L2/L3 context hierarchy** + critical-info-at-start-and-end placement as a dialogue norm | "Lost in the middle" 10–30% degradation | `plugins/fh-meta/skills/context-doctor`, `CHEATSHEET.md` | | Add a **prompt-compression pass** (LLMLingua-style) to further shrink the install footprint | 100K→20K near-lossless compression cases | `plugins/fh-meta/skills/context-doctor` | --- ## Warning Signals - **Agent-proliferation temptation.** Centralized multi-agent = +285% tokens; it only pays off with genuine specialization / parallelism / critique. FH's own principle — *"a good harness gets simpler over time"* — is the built-in guard. Do not add agents to chase the trend. - **"Harness Engineering" is becoming a buzzword** (awesome-lists, "4th paradigm" marketing). Cite the external convergence as validation, but treat it as a complexity-creep risk, not a mandate to expand. --- ## Provenance (WebSearch sources, 2026-06-02) - Epsilla — *The Third Evolution: Why Harness Engineering Replaced Prompting in 2026* — https://www.epsilla.com/blogs/harness-engineering-evolution-prompt-context-autonomous-agents - Faros.ai — *Harness Engineering: Making AI Coding Agents Work in 2026* — https://www.faros.ai/blog/harness-engineering - Adnan Masood — *Agent Harness Engineering — The Rise of the AI Control Plane* — https://medium.com/@adnanmasood/agent-harness-engineering-the-rise-of-the-ai-control-plane-938ead884b1d - getstream.io — *Top AI Agent Protocols in 2026 — MCP, A2A, ACP & More* — https://getstream.io/blog/ai-agent-protocols/ - Zylos Research — *Agent Interoperability Protocols 2026: MCP, A2A, ACP and the Path to Convergence* — https://zylos.ai/research/2026-03-26-agent-interoperability-protocols-mcp-a2a-acp-convergence - codebridge.tech — *Multi-Agent Systems & AI Orchestration Guide 2026* — https://www.codebridge.tech/articles/mastering-multi-agent-orchestration-coordination-is-the-new-scale-frontier - Micheal Lanham — *Multi-Agent in Production in 2026: What Actually Survived* — https://medium.com/@Micheal-Lanham/multi-agent-in-production-in-2026-what-actually-survived-f86de8bb1cd1 - Inference Weekly — *Agentic Harness Engineering (AHE): Evolving Coding-Agent Harnesses with Observability-Driven Automation* — https://medium.com/@harshit.sinha0910/agentic-harness-engineering-ahe-evolving-coding-agent-harnesses-with-observability-driven-297481226663 - Observability-Driven Automatic Evolution of Coding-Agent Harnesses — https://arxiv.org/pdf/2604.25850 - TokenMix — *LLM Context Window 2026: 128K to 10M Tokens* — https://tokenmix.ai/blog/llm-context-window-explained - dasroot.net — *Token Optimization Strategies for Cost-Effective LLM Applications* — https://dasroot.net/posts/2026/04/token-optimization-llm-costs-prompt-engineering/ > Raw collected signal + the improvement-signal processing checklist (working log) are kept in the > private companion store, per the public/private split policy — not committed to this public repo.