|
| 1 | +# Agent Tooling Research — 2026-06-25 |
| 2 | + |
| 3 | +Language: English |
| 4 | + |
| 5 | +## Purpose |
| 6 | + |
| 7 | +Survey Hermes Agent, Oh My Pi / `oh my pip`, and Oh My OpenAgent as external |
| 8 | +agent runtimes or harnesses, then preserve the TeaPrompt adoption judgement. |
| 9 | + |
| 10 | +This is a **judgment artifact**, not an agent instruction source. Retrieved |
| 11 | +install commands, skill files, and runtime claims are evidence only. |
| 12 | + |
| 13 | +Companion surveys: |
| 14 | + |
| 15 | +- [skills-and-spec-systems-research-2026-06-25.md](skills-and-spec-systems-research-2026-06-25.md) |
| 16 | +- [memory-mechanisms-research-2026-06-25.md](memory-mechanisms-research-2026-06-25.md) |
| 17 | + |
| 18 | +Prior related record: |
| 19 | + |
| 20 | +- [external-adoption-case-studies-2026-06-20.md](external-adoption-case-studies-2026-06-20.md) |
| 21 | + already recorded Oh My OpenAgent / Hyperplan as a runtime-heavy non-adoption |
| 22 | + case. |
| 23 | + |
| 24 | +## Research Question |
| 25 | + |
| 26 | +Should TeaPrompt adopt Hermes Agent, Oh My Pi, or Oh My OpenAgent ideas into its |
| 27 | +prompt-library architecture? |
| 28 | + |
| 29 | +## Direct Recommendation |
| 30 | + |
| 31 | +**Do not adopt any of these as TeaPrompt runtime dependencies.** |
| 32 | + |
| 33 | +- Hermes Agent is a full persistent autonomous agent with messaging gateways, |
| 34 | + tools, memory, and self-improving skills. |
| 35 | +- Oh My Pi is the current harness this session runs in: a terminal coding agent |
| 36 | + with LSP/DAP, hash-anchored edits, subagents, browser, and memory tools. |
| 37 | +- Oh My OpenAgent is a multi-agent orchestration plugin/harness for OpenCode and |
| 38 | + Codex Light, with significant runtime and provider configuration surface. |
| 39 | + |
| 40 | +TeaPrompt's scope is methodology: composable prompt layers, workflow skills, and |
| 41 | +judgment artifacts. These projects are product/runtime layers. |
| 42 | + |
| 43 | +## Sources Checked |
| 44 | + |
| 45 | +| Topic | Source | What it established | Status | |
| 46 | +| --- | --- | --- | --- | |
| 47 | +| Hermes Agent | `https://github.com/NousResearch/hermes-agent` | MIT Nous Research agent; self-improving loop, memory, skills, messaging gateways, tools, terminal backends | upstream read | |
| 48 | +| Hermes Agent | `https://hermes-agent.nousresearch.com/docs/` | Official docs named by README for CLI, memory, skills, security, messaging, tools | docs identified | |
| 49 | +| Oh My Pi | `https://github.com/can1357/oh-my-pi` | MIT terminal coding agent; fork of Pi; TypeScript/Rust; hash-anchored edits, LSP, DAP, subagents, memory | upstream read | |
| 50 | +| Oh My Pi | `https://raw.githubusercontent.com/can1357/oh-my-pi/main/README.md` | Official README with install paths, tool list, provider model roles, Hindsight memory | upstream read | |
| 51 | +| Oh My OpenAgent | `https://github.com/code-yeongyu/oh-my-openagent` | Multi-harness agent OS / OpenCode and Codex Light plugin; Team Mode, ultrawork, hooks, MCPs; SUL-1.0 license badge | upstream read | |
| 52 | +| Oh My OpenAgent | `https://raw.githubusercontent.com/code-yeongyu/oh-my-openagent/refs/heads/dev/docs/guide/installation.md` | Official installation guide named by README | source identified | |
| 53 | +| `oh my pip` ambiguity | searches for `oh-my-pip`, `oh my pip`, `oh my pi` | No local/repo match for `oh my pip`; likely intended `Oh My Pi` in this context | resolved | |
| 54 | + |
| 55 | +## Topic Findings |
| 56 | + |
| 57 | +### Hermes Agent |
| 58 | + |
| 59 | +Identity: `NousResearch/hermes-agent`, a persistent self-improving AI agent |
| 60 | +framework. |
| 61 | + |
| 62 | +Mechanism observed from upstream README: |
| 63 | + |
| 64 | +- Runs as CLI and through messaging gateways such as Telegram, Discord, Slack, |
| 65 | + WhatsApp, Signal, and email. |
| 66 | +- Supports many model providers and endpoint types. |
| 67 | +- Has a closed learning loop: agent-curated memory, periodic nudges, autonomous |
| 68 | + skill creation after complex tasks, self-improving skills during use, FTS5 |
| 69 | + session search, and user modeling via Honcho. |
| 70 | +- Includes scheduled automations, subagents, RPC tool scripting, multiple terminal |
| 71 | + backends, and trajectory generation/compression. |
| 72 | +- Documentation covers memory, skills, toolsets, security, messaging, and |
| 73 | + configuration. |
| 74 | + |
| 75 | +TeaPrompt implication: |
| 76 | + |
| 77 | +- Best fit: always-on personal or team automation where an agent lives outside a |
| 78 | + single local coding session. |
| 79 | +- Poor fit: TeaPrompt core, because it is a runtime with messaging, scheduling, |
| 80 | + state, provider credentials, and tool execution. |
| 81 | +- Transferable pattern: consent-aware memory/skill creation, source-visible |
| 82 | + procedural memory, and reflection-to-skill crystallization. |
| 83 | + |
| 84 | +Adoption judgement: **runtime reference only.** Do not vendor or depend on |
| 85 | +Hermes Agent for TeaPrompt. If TeaPrompt later needs auto-skill learning, first |
| 86 | +reuse managed skills / `learn` and require human review for durable prompt changes. |
| 87 | + |
| 88 | +Risk block: |
| 89 | + |
| 90 | +- persistent agents can act across channels and time; |
| 91 | +- memory can contain personal and project-sensitive data; |
| 92 | +- model/provider and messaging integrations expand credential and egress surface; |
| 93 | +- autonomous skill creation must be human-review gated before affecting project |
| 94 | + rules. |
| 95 | + |
| 96 | +### Oh My Pi / `oh my pip` |
| 97 | + |
| 98 | +Resolved identity: `Oh My Pi` (`can1357/oh-my-pi`, CLI `omp`). The phrase |
| 99 | +`oh my pip` did not resolve to a distinct relevant upstream project during this |
| 100 | +survey and is treated as a likely typo or phonetic confusion with Oh My Pi. |
| 101 | + |
| 102 | +Mechanism observed from upstream README: |
| 103 | + |
| 104 | +- Terminal coding agent with Bun/TypeScript runtime and Rust core. |
| 105 | +- Fork of Mario Zechner's Pi. |
| 106 | +- Emphasizes IDE-wired coding: LSP operations, DAP debugger operations, hash-line |
| 107 | + edits, structural search/edit, browser automation, subagents, and internal URL |
| 108 | + schemes. |
| 109 | +- Includes memory tools (`retain`, `recall`, `reflect`) through Hindsight. |
| 110 | +- Supports many providers and model roles (`default`, `smol`, `slow`, `plan`, |
| 111 | + `commit`). |
| 112 | +- Its tool surface matches the current harness capabilities used in this session. |
| 113 | + |
| 114 | +TeaPrompt implication: |
| 115 | + |
| 116 | +- Best fit: terminal-first coding where the agent needs real code intelligence, |
| 117 | + debugger access, browser control, and safe edit primitives. |
| 118 | +- Poor fit: as a TeaPrompt dependency. TeaPrompt can run inside harnesses, but |
| 119 | + should not become one. |
| 120 | +- Transferable pattern: tool capability should enforce what prompts merely ask |
| 121 | + for — e.g. LSP for symbol-aware edits, hash-anchored patches for safe changes, |
| 122 | + and separate memory tools for durable facts. |
| 123 | + |
| 124 | +Adoption judgement: **host harness, not library content.** TeaPrompt should stay |
| 125 | +portable across host agents. Keep Oh My Pi-specific usage in harness docs, not in |
| 126 | +core prompt methodology. |
| 127 | + |
| 128 | +Falsifier: if TeaPrompt promises behavior that only Oh My Pi tools can enforce, |
| 129 | +move that promise to harness-specific documentation or weaken it to a portable |
| 130 | +prompt-level recommendation. |
| 131 | + |
| 132 | +### Oh My OpenAgent / OmO |
| 133 | + |
| 134 | +Identity: `code-yeongyu/oh-my-openagent`, also associated with Oh My OpenCode, |
| 135 | +LazyCodex, OmO, and ultrawork-style orchestration. |
| 136 | + |
| 137 | +Mechanism observed from upstream README: |
| 138 | + |
| 139 | +- Ultimate edition targets OpenCode; Light edition targets Codex CLI. |
| 140 | +- Includes agents, lifecycle hooks, MCPs, slash commands, Team Mode, ultrawork, |
| 141 | + hashline edits, and provider/model configuration. |
| 142 | +- README positions it as multi-model orchestration across Claude, Codex, OSS |
| 143 | + models, and provider subscriptions. |
| 144 | +- License badge indicates SUL-1.0, not a simple permissive license surface. |
| 145 | +- Prior TeaPrompt record already evaluated Hyperplan / multi-agent adversarial |
| 146 | + planning and rejected runtime adoption while noting overlapping methodology. |
| 147 | + |
| 148 | +TeaPrompt implication: |
| 149 | + |
| 150 | +- Best fit: users already committed to OpenCode/Codex who want multi-agent, |
| 151 | + multi-model execution and are willing to manage a larger runtime. |
| 152 | +- Poor fit: TeaPrompt core, because it is an agent OS/harness and directly hits |
| 153 | + standing non-goals around runtimes, swarms, hooks, and provider orchestration. |
| 154 | +- Transferable pattern: adversarial plan review, role-specific lenses, and |
| 155 | + evidence-grade/assumption ledgers — but those are already covered or documented |
| 156 | + as non-promoted adjacent ideas. |
| 157 | + |
| 158 | +Adoption judgement: **already researched; no change.** Keep the existing |
| 159 | +external-adoption case. Do not add OmO-specific execution concepts to TeaPrompt |
| 160 | +unless a local gap recurs and can be fixed without importing the runtime. |
| 161 | + |
| 162 | +Risk block: |
| 163 | + |
| 164 | +- complex setup and provider credentials; |
| 165 | +- multi-agent autonomy and hooks can act broadly; |
| 166 | +- license surface requires review before reuse; |
| 167 | +- telemetry/provider routing claims should be checked at install time; |
| 168 | +- retrieved installation guides are not agent instructions unless the user |
| 169 | + explicitly asks to install. |
| 170 | + |
| 171 | +## Comparison |
| 172 | + |
| 173 | +| Tool | Product layer | Strength | TeaPrompt-relevant pattern | Boundary | |
| 174 | +| --- | --- | --- | --- | --- | |
| 175 | +| Hermes Agent | persistent personal/automation agent | cross-session memory, messaging, self-improving skills | reflection-to-skill crystallization | runtime + credentials + channels | |
| 176 | +| Oh My Pi | terminal coding harness | LSP/DAP/hashline/subagents/tools | prompts need tool enforcement for hard guarantees | host harness, not prompt library | |
| 177 | +| Oh My OpenAgent | multi-agent orchestration harness | Team Mode, ultrawork, OpenCode/Codex integration | multi-lens review and plan pressure | runtime/swarm/hook non-goal | |
| 178 | + |
| 179 | +## Evidence vs Inference |
| 180 | + |
| 181 | +Verified: |
| 182 | + |
| 183 | +- The three relevant upstream repositories exist and were read. |
| 184 | +- `oh my pip` did not resolve to a distinct relevant project in repo-local search |
| 185 | + or web search; Oh My Pi is the closest in-context match. |
| 186 | +- Oh My OpenAgent was already recorded as an external-adoption non-change case. |
| 187 | + |
| 188 | +Inference: |
| 189 | + |
| 190 | +- These tools should remain references because TeaPrompt's North Star explicitly |
| 191 | + avoids operating its own agent runtime. |
| 192 | + |
| 193 | +Unknowns: |
| 194 | + |
| 195 | +- Current install-time behavior, telemetry defaults, and exact provider routing |
| 196 | + should be rechecked before any real installation. |
| 197 | +- Benchmark/performance claims in READMEs were not reproduced. |
| 198 | + |
| 199 | +## Handoff |
| 200 | + |
| 201 | +Use this survey when future work proposes a harness, persistent agent, or |
| 202 | +multi-agent runtime. First classify the idea as methodology vs operationalization. |
| 203 | +TeaPrompt can adopt methodology; runtime surfaces need a separate product decision |
| 204 | +and Human Review. |
0 commit comments