Abhed — Build Plan, Team & Sequencing

Status: Draft · 2026-09-02 Evidence status: [E]. Research produced no verified claims on build cost, team, or timeline. Estimates below are engineering judgment, calibrated to comparable platform work.

1. Build vs adopt

The principle: build the harness, adopt everything else. The harness is where the 7.80× variance lives (P1) — it is the only layer where custom engineering returns more than it costs. Everything below it is undifferentiated infrastructure.

Reference implementation worth studying: BeeAI Framework is Apache-2.0, Linux Foundation-hosted (AI & Data, incubation status), actively maintained, with a provider-agnostic provider:model backend abstraction and 13 documented providers. Apache-2.0 means forking for air-gapped use is legally clean.

Two caveats, both verified: its own README states the code is provided with no support commitment and that it "will not be maintaining this code going forward"; and its Backend abstracts mostly hosted provider SDKs, so on-prem paths are the minority. Study the abstraction, don't inherit the dependency.

License discipline: prefer Apache-2.0/MIT throughout. Avoid AGPL in anything linked into the product, and avoid source-available licenses with field-of-use restrictions entirely — they are incompatible with shipping Abhed to customers.

2. Phasing

Phase 0 — Foundations (weeks 1–6)

Prove the loop and the economics before building a platform around them.

• Event store + step(state) loop with all terminal events (P6, arch §2)
• Model adapter + capability probe against one OpenAI-compatible endpoint
• Native tool set: read, write, edit, glob, grep, bash
• Policy engine with ordered evaluation and absolute deny (P7)
• CLI (Go static binary — matters for air-gap install)
• Prefix-cache measurement harness — validate the 17× assumption on real hardware

Exit criteria: agent completes a multi-file change on a trusted repo; every terminal path exercised; prefix-cache hit rate measured and > 70%.

Phase 1 — Harness depth (weeks 7–16)

This is where agent quality is actually won.

• Compaction with PreCompact hook + ABHED.md re-injection (P4)
• Subagent supervisor with hierarchical budget enforcement (P3)
• JIT retrieval tier 0/1 (grep/glob + tree-sitter repo map)
• Eval harness: task suite plus automated log inspection (P10)
• Composable loop strategies: ReAct spine + generate-test-repair (P5)

Exit criteria: measurable improvement on your own eval suite across ≥2 model families; log inspection catches at least one class of misbehavior that scores miss.

Phase 2 — Platform (weeks 17–28)

• Server mode: multi-user, tenancy, OIDC SSO, managed settings
• Web console (session view, approvals, audit, replay)
• Firecracker execution pool with per-session isolation
• MCP gateway + internal registry with review gate
• OpenTelemetry tracing end to end

Exit criteria: two tenants isolated and verified; security checklist (§03-security §7) passing; audit replay reconstructs a session exactly.

Phase 3 — Air-gap & scale (weeks 29–40)

• Signed offline bundle + installer; digest-pinned everything
• Egress broker with allowlist and full audit
• Retrieval tier 2 (hybrid + rerank) for monorepo scale
• Multi-node inference; capacity model validated against real load
• Compliance evidence package

Exit criteria: clean install into a network-isolated enclave from bundle only; zero egress verified by test.

Phase 4 — Differentiation (ongoing)

• Cross-model consistency benchmark (P12) — the asset no one else has
• Declarative requirement/rule constraints, measured not assumed
• Per-model, per-task-class effort policy (P11)
• Continuous harness A/B against the eval suite

3. Team

The eval engineer is not optional. Given P1, a team that cannot measure harness changes is flying blind on the single variable that matters most. Hire that role in Phase 0, not Phase 3.

4. Effort estimate [E]

At fully-loaded cost this is roughly $1.5–2.5M of engineering — likely more than the T2 hardware it runs on. That ratio is normal and worth stating plainly to sponsors: the GPUs are the visible cost, the harness is the real one.

A credible demo is reachable in 6 weeks (Phase 0). A credible product is not.

5. Risk register

6. Open questions to resolve with your own data

Carried directly from the research pass — these are unresolved in the literature, and answering them on your own hardware is genuine contribution:

1. Does the 7.80× harness-variance finding hold for self-hosted open-weight models? The factorial used frontier models. If harness variance is larger for weaker on-prem models, harness investment pays off more on-prem. If it's driven by frontier models' ability to exploit rich scaffolds, the conclusion inverts and thin-harness/strong-model becomes correct. This determines Abhed's entire engineering budget split. Testable.
2. What are the GPU-second economics of subagent fan-out? Nobody has published tokens-per-solved-task mapped onto GPU-seconds, nor the crossover where a subagent's 1–2k summary stops repaying its own prefill.
3. How does prefix caching behave under compaction and subagent fan-out in a real serving engine? Phase 0 answers this.
4. Does declarative rule-constraining actually reduce cross-model variance, and by how much? Zero independent validation exists. Measuring it is Abhed's differentiator.
5. Is there any sandboxing/injection-defense evidence that survives scrutiny? Next research target; currently a blind spot.