Reading the agent landscape

01 · Managed Agents vs Valor

Valor orchestrates, Managed Agents executes

Claude Managed Agents is Anthropic's fully hosted agent harness — a pre-built agent loop, managed cloud containers, and session persistence, all driven over an API. The study's governing thought is a division of labour: adopt Managed Agents for the hands, keep the brain local. Three layers of Valor stay untouched by any hosted engine — orchestration (the PM/Dev session split and the SDLC pipeline), integration (Telegram and email delivery, a dozen external services), and memory (ranked, decaying recall).

Where the two meet is execution. The BUILD and TEST stages are the clear offload candidates: they are the longest-running, the most self-contained, and the ones that benefit most from a clean sandboxed container instead of the local worktree machinery. The remaining stages stay local because they need Telegram context, git credentials, or the review gate's own eyes.

The hard problem is subconscious memory. Valor injects <thought> blocks mid-turn through Claude Code hooks; Managed Agents has no hook equivalent and no way to insert tokens into a generation in progress. The study's workaround is stealth injection through the workflow tools the agent must call anyway — run_tests, git_commit, submit_work — each routed through Valor so its output can carry recalled memory the agent never asked for.

The conclusion is a trigger, not a migration. When Managed Agents ships production memory alongside a hook-equivalent API, the cost of maintaining custom execution infrastructure flips. Until both land, the memory system is the moat, and the architecture stays hybrid.

02 · Valor vs Paperclip

A vertical teammate meets a horizontal control plane

Paperclip is an open-source control plane for orchestrating teams of agents into structured organizations: org charts, per-agent budgets, board-level approval gates, and goal-aligned task decomposition. Its metaphor is "a company you are managing, not a tool you are using." Valor is the opposite shape — a vertically integrated development teammate that owns every stage from a Telegram message to a merged pull request, and knows exactly what an SDLC pipeline and a PR review are.

The study frames them as complementary rather than competitive. Paperclip is a horizontal layer that coordinates any agents doing any work; it has no memory system and no opinion about software. On the weighted decision matrix for Valor's single-project use case the score is 7.2 to 4.2 — but Paperclip wins decisively on the two axes Valor is weakest at: budget governance and multi-agent breadth.

The recommendation is to adopt Paperclip's best ideas natively before adopting Paperclip itself. Per-session and per-project token budget hard-stops, goal-aligned tracing from session to issue to objective, and an immutable append-only audit log all fit inside Valor's existing PM orchestration without a new dependency. Integrating the platform becomes worth its engineering cost only if Valor scales past three concurrent projects with multiple agent types.

03 · Managed Agents vs Perplexity Computer

A programmable engine, not a digital worker

This study puts two execution backends side by side and finds a fork rather than a spectrum. Managed Agents is API-first and developer-oriented: a Linux container the agent works inside, created, streamed, steered, and archived over a REST interface. Perplexity Computer is UI-first and end-user-oriented: an autonomous digital worker that interprets rendered web pages with a vision-language model and clicks, types, and scrolls the way a person would.

For a system whose whole job is to orchestrate execution programmatically, that difference settles it. Managed Agents exposes the full lifecycle an orchestrator needs; the Computer product exposes none of it — there is no documented API to create tasks, stream progress, inject steering, or harvest structured results. On Valor's weighted criteria the split is 8.9 to 3.7, with Managed Agents winning on programmatic orchestration, code execution, and cost predictability — per-millisecond billing against an opaque, complexity-driven credit system where a single codebase scan once burned twice a month's allotment.

Perplexity Computer's one real advantage is its visual browser, which handles dynamic and paywalled pages that a text fetch cannot. The recommendation follows the strengths: adopt Managed Agents as the execution backend, evaluate Perplexity's separate Agent API as a supplementary research tool for the planning and critique stages, and do not try to bend the Computer product into a backend it was never designed to be.

04 · OpenHuman vs Hermes

Memory shape follows form factor

The fourth study steps outside Valor to compare two self-hosted agent stacks that landed in the same window and made very different bets. OpenHuman is a desktop app for one person's digital life: a deterministic memory tree that pulls from a hundred-plus OAuth sources, normalizes everything to provenance-tagged Markdown you can read and edit by hand, and folds it into browsable summary trees. It bets on transparency over autonomy — no vector-soup black box, and the audit trail is just files on disk.

Hermes is the inverse: a headless framework you deploy on a cheap VPS and pipe through twenty chat platforms, with agent-curated memory, first-class self-improving skills, and six interchangeable execution backends. It bets on autonomy and reach over transparency — memory is something the agent earns, and "what does it know" becomes a query rather than a folder.

The lessons generalize past either product. Memory shape follows form factor: a single-user desktop app can afford a Markdown vault, while a headless multi-tenant agent needs an indexed session store with lineage across compressions. Sandboxing is a deployment choice, not a feature — Hermes ships six backends because operators need the range, OpenHuman ships one because there is one operator. And self-improving, shareable skills remain the live experiment worth watching whether or not you build on them.