Research Epic Manager | Agent Skills

Autonomous research as an agent skill: This repo contains a collection of agent skills which enable an agent to run scientific research projects, from idea to manuscript.

In order to organise the long-horizon work of a research project, we use an Agile-inspired tracking system. Plain markdown files (EPIC.md, ISSUE.md, comments/*.md) track all work, keeping orchestration and revision documents separate from project files. Agent skills encode the bookkeeping of this system, operating completely within the local file system. Human feedback is integrated naturally, from agent conversations → issue comments.

The core "tracker" skills are highly flexible and are useful for organising any long-horizon project. Hence, the "research project management" component is defined as a specialised "epic manager" skill — to define the lifecycle of what issues need to be generated throughout the project from idea, development and revisions.

Installation

DeepSeek Harness (DSH)

1. Symlink or copy skills/ into your discovery root (.dsh/skills under your workspace root).
2. Paste AGENTS.md into your tracker root (EPICS_ROOT, default /workspace/epics/); the copy inside skills/ stays with the skill directory, and the root copy is what the harness reads as workspace instructions. If your layout differs, edit the two path definitions in skills/AGENTS.md.

Other harnesses

These agent skills are not harness-agnostic. Currently, only DSH has been tested.

That being said, there is actually no hard dependency on DSH. To install in your harness:

1. Copy skills/ and AGENTS.md into your harness-specific locations.
2. Overwrite the DSH-native sub-agent dispatch mechanics in issue-subagent-orchestration and research-project-epic-manager skills for your harness-specific conventions.

Skill catalog

Skill	One-liner
add-epic	Create an epic container (directory + EPIC.md)
add-issue	Add an issue (draft-first lifecycle + readiness gate → open)
add-comment	Append an anonymous, chronological comment
issue-manager	Protocol for a chat working an issue (roles, dispatch, verification)
issue-subagent-orchestration	DSH-native: documents the harness's delegation tools (subagent, workflow, send_message, job_output, interrupt_agent)
research-project-epic-manager	Specialised epic manager (DSH-native dispatch mechanics): spins up issues + issue managers to take projects through a 5-phase Scope → Generate → Dispatch → Collect → Synthesize loop, with gated verdicts

DeepSeek-Harness-specific bits: skills/issue-subagent-orchestration/SKILL.md is harness-native: it documents the harness's delegation tools (subagent, workflow, send_message, job_output, interrupt_agent). The research-project-epic-manager skill's dispatch mechanics (the Phase-C brief templates in its references/) likewise reference the harness's delegation tools (subagent, job_output, send_message, interrupt_agent) and must be mapped onto another runtime's delegation tools.

User guide

Start a whole project by asking the chat to "manage a research project":

Manage a research project - create a workshop paper under https://www.climatechange.ai/events/neurips2026 as the theme. Limitations: CPU-only data analysis. Deliverable: a 4-page PDF with workshop formatting.

The orchestrator takes a research brief and a project config (start from skills/research-project-epic-manager/assets/project-config-template.yaml) and runs five phases: scoping, just-in-time issue generation per stage, dispatch of coder/reviewer children, staging/collecting of stage outputs into a flat project.json manifest, and synthesis of the deliverable.

Watching it: everything is files — EPIC.md (issue index), project.json (manifest), the append-only verdict_history on the orchestration log, and the project workspace plus the project.json manifest's append-only artifacts map. Any "status of epic <X>?" works.

You, mid- or post-run: in the default autonomous mode the run does not stop for approval — it decides, records the decision, and proceeds. Your steer lands as an anonymous [human-directive:<digest>] comment on the affected issue, dated and routed, appended before it is acted on; project-level input also lands on the orchestration log. Prefer approval points? Set autonomy.mode: gated in the config. Either way, genuine blockers (exhausted review loops, catastrophic failures, missing route targets) surface as blocked issues with manager notices; unblock them with a directive or by working the issue yourself ("work on issue <X>" makes that chat the issue manager, who verifies before closing anything).

Act as issue manager for <path to issue>: ... describe a request/comment/question

Or talk directly to the epic manager:

Include a constraint to this project ...

Comparison to other systems

The table grades the 2024–2026 cohort of autonomous research systems.

How to read the marks

Mark	Meaning
✅	yes — demonstrated in the surveyed sources
🟡	partial — some form exists, or a claim is not independently verified
❌	no — nothing found
❔	not recorded — absence of evidence (not evidence of absence)

Column scales: Setup (entry requirement you face): 📖 open source · 🔒 closed · 🧪 paper-only prototype · ❔ not recorded • Lifecycle coverage (phases that run unaided): ✅ full loop · 🟡 decisive subset · ❌ single phase • Lifecycle auditability (can you reconstruct how the run proceeded): ✅ decision record · 🟡 artifacts only · ❌ nothing inspectable · ❔ not recorded • Self-improvement (does the system improve itself, weakest→strongest): 🔄 object-level · 🧠 memory-informed · 🎓 weight-level · 🪞 protocol-level.

System	Setup	Lifecycle coverage	Lifecycle auditability	Self-improvement
research-epic-manager (this repo)	📖 six SKILL.md files + stdlib batteries; Python 3 only	✅ full loop — scoping → synthesis, gates included	✅ decision record — append-only verdict history, executed backward invalidation	🪞 protocol-level — gate failures become findings; fixes tracked in-tracker
AI-Scientist v1/v2	📖 OSS repo + LLM keys + GPU	✅ full loop — ideation → experiment → write-up → LLM review	🟡 artifacts only — run dirs + cost tracking	🔄 object-level — tree search within one paper
CycleResearcher	📖 open-weight models	🟡 writing + review half	🟡 artifacts only — training-loop records	🎓 weights — retrains its own models
Agent Laboratory / AgentRxiv	📖 OSS; frontier-model dependent	🟡 lit review → experiment → report	🟡 artifacts only — feedback lives in conversation	🔄 object-level — co-pilot steering
STORM / Co-STORM	📖 OSS pip package	❌ pre-writing only	🟡 artifacts only — mind map persists	❌ none recorded
DeepScientist	❔ containerised; month-scale	🟡 hypothesise → verify → analyse tiers	🟡 artifacts only — Findings Memory; forward-only	🧠 memory — failures retained across runs
Kosmos	❔ not recorded	🟡 long-horizon discovery; write-up unclear	🟡 artifacts only — world model summaries each cycle	🧠 memory — world model steers next cycle
ASI-Arch	❔ not recorded — ~20k GPU-hours	❌ architecture discovery only	🟡 artifacts only — every trial archived, searchable	🧠 memory — trial archive feeds future cycles
CodeScientist	🧪 release not recorded	🟡 ideation → construct → report	🟡 artifacts only — judge-scored stages	❔ not recorded
Co-Scientist	🔒 limited access	🟡 hypothesis + protocol design	🟡 artifacts only — debate records + Elo	🔄 object-level — tournament evolution
Robin	🔒 proprietary	🟡 propose → analyse loop; humans run the lab	❔ not recorded	🔄 object-level — re-propose iterations
Zochi	🔒 closed, doc-only internals	🟡 claimed full loop — unverified	❌ nothing inspectable	❔ not recorded
FunSearch	🔒 closed — Nature paper	❌ search-discovery only	❔ not recorded	🔄 object-level — evolving program islands
ResearchAgent	🧪 no public release	❌ ideation only	🟡 artifacts only — ~5 scored rounds described	🔄 object-level — reviewer-scored refinement

Validation

Sanity check: ask the LLM to run the validation batteries (each skill's scripts/test_*.py) and expect exit code 0; or ask the LLM to explain how the system works and check the explanation against the docs. The batteries are stdlib-only Python and self-deleting: test_add_epic.py, test_add_issue.py and test_add_comment.py drive the real skill scripts against a scratch epic that removes itself on exit (unless --keep), and test_lifecycle.py additionally enforces cross-cutting invariants (discovery roots, git hygiene, layout, and the ASCII-only rule for .md files). The research-project-epic-manager skill's own batteries are scripts/validate_execution.py and scripts/validate_protocol.py.

Contributing

See CONTRIBUTING.md: MIT license, collective field-level attribution, anonymous tracker comments vs. normal git history, the ASCII-only rule for agent-facing files, and how to validate a change.