Inside our Managed Agents architecture — 15 agents, 1 environment

When Anthropic shipped Managed Agents, the obvious-but-wrong move was to take our existing inline cascade and lift each call into a separate agent. That way lies $3.50-per-report territory. Instead we treated the platform as a chance to redraw the system around user actions first, with specialists called only where the domain work is non-trivial.

The result: 15 agents in one environment, mapped 1:1 to product surfaces, with a thin HMAC-signed bridge sitting between them and our PostgreSQL writes.

Primary layer — seven user-facing agents

Each primary agent maps to exactly one product action. No more, no less:

• P1 Main Orchestrator (Haiku 4.5) — text-router for ambiguous chat input. Cheap, fast, almost always right.
• P2 Area Analysis Pipeline (Opus 4.7) — the heavy hitter. Drawn polygon in, nine-section report out, fanning to four specialists.
• P3 Report Translator (Sonnet 4.6) — section-parallel EN ↔ TR/RU/UZ translation. Sibling report, glossary-pinned.
• P4 Section Re-analyzer — single-section regenerate with optional user hint, model picked at the call site.
• P5 Oblast Aggregator (Opus 4.7) — N parcels in, viloyat-level brief out, parcel-weighted statistics.
• P6 Chat Companion (Sonnet 4.6) — mascot chat backed by our public MCP server. No fabrication; everything grounded in MCP tools.
• P7 Share & PDF Orchestrator (Haiku 4.5) — mints share tokens, drives PDF generation, crafts OpenGraph preview text.

Specialist layer — eight domain workers

Specialists are JSON-in / JSON-out. They never address the user; they get a structured task, return structured data, and the primary stitches the result.

• S1 Cadastral Analyst (Sonnet 4.6) — spatial → structured cadastral JSON.
• S2 Satellite Interpreter (Opus 4.7) — 12-month NDVI history → vegetation-health JSON.
• S3 Policy Aligner (Sonnet 4.6) — analysis + WB / UZ-2030 / UNFF / SDG framework docs → alignment JSON.
• S4 Composer (Opus 4.7) — S1+S2+S3 outputs → nine XML-tagged markdown sections.
• S5 Species Recommender (Sonnet 4.6) — bioclimate + soil + 64-species catalog → ranked 3–5 recommendations.
• S6 Cost Estimator (Haiku 4.5) — species + area + oblast → USD cost bands with multipliers.
• S7 Risk Analyzer (Sonnet 4.6) — spatial + climate + infrastructure → categorised risks with mitigations.
• S8 Translator (Sonnet 4.6) — section markdown + target lang → translated section markdown.

The bridge that ties it together

Agents live in Anthropic's environment. Our database lives on our VPS. The bridge — five small modules in backend/app/ — handles every crossing:

• auth/agent_token.py — HMAC-signed 60-second service-to-service tokens. An agent's reply is rejected if the signature is missing, replayed, or mismatched.
• models/agent_invocation.py — every call logged with cost, latency, tokens, status. Full audit trail.
• routes/agent_internal.py — HMAC-protected /api/agent/internal/* endpoints (get_spatial_context, write_section, finalize_report). Internet can't reach them.
• agent_bridge/* — SSE event translator that streams agent progress to the React frontend.
• Contract tests — 9/9 green for the P2 path.

Why the production flag is OFF

USE_AGENT_PIPELINE=false by default. The Managed Agents path costs roughly 3× the inline cascade because each sub-agent rebuilds part of its context. For an always-on demo deployment, the inline path is the right baseline.

Where we'll flip it ON: cross-team workflows, the researcher sub-agent builder on the Q3 2026 roadmap, and parallel report queues for multi-tenant orgs. Anywhere isolation matters more than the bill.

What we'd ship next

• Per-agent budget caps. Right now the cost guardrail is a daily total ($50 alert, $100 hard-stop). Per-agent caps would let us experiment with Opus-on-everything safely.
• Agent-level prompt cache namespacing. S1, S2, S3 share most of their grounding context — caching them in a single namespace would shave another 30%.
• A "shadow" run mode. Send the same request through both inline and Managed Agents paths, diff the outputs, surface where the architectures disagree. Best ground-truth we can build for ourselves.

Architecture diagram + per-agent matrices live in docs/HACKATHON_AGENT_ARCHITECTURE.md in the repository.